April 8, 2026 • 38 mins
Your AI can write code fast, but it can also wander fast. That’s why we sat down with Jason Belk from Learning at Cisco to unpack spec-driven development, a simple idea with huge impact: write the rules and requirements first, then let your coding agent execute with far fewer surprises.
We talk through what “agentic coding” looks like in practice with Claude Code, including the trust and permission model of a local AI agent that can create files, run bash commands, and iterate on a real project. Jason explains how GitHub Spec Kit turns plain markdown and scripts into a repeatable workflow: start with a constitution that defines governing principles, then cycle feature by feature through specify, plan, tasks, and implement. Along the way we cover common gotchas like initializing in the wrong directory so skills never load, plus practical tips like using voice-to-text to improve prompts and choosing the right model tier when implementation quality matters.
We also zoom out to the bigger picture: why context windows break long builds, how keeping plans on disk helps the agent “re-ground” itself, and where the industry may be heading with small specialized models versus one giant general LLM. Jason shares learning resources too, including a Cisco U tutorial that frames spec-driven development for network engineers, and the Cisco AI Technical Practitioner course and certification, plus upcoming Cisco Live sessions.
Subscribe for more real-world AI workflows, share this with a teammate who keeps fighting prompt drift, and leave a review with the one automation project you want an agent to build next.
Connect with Jason:
https://linktr.ee/renobelk
https://github.com/jabelk/claude-speckit-template
SDD Relevant Material mentioned in this episode:
https://substack.com/home/post/p-189415335
https://ondemandelearning.cisco.com/apollo-alpha/tc-ai-spec-driven-dev/pages/1 https://learningnetwork.cisco.com/s/aitech-exam-topics
https://u.cisco.com/paths/cisco-ai-technical-practitioner-20806
Software Defined Talk episode referenced:
https://www.softwaredefinedtalk.com/563
Check out the Monthly Cloud Networking News
https://docs.google.com/document/d/1fkBWCGwXDUX9OfZ9_MvSVup8tJJzJeqrauaE6VPT2b0/
Visit our website and subscribe: https://www.cables2clouds.com/
Follow us on BlueSky: https://bsky.app/profile/cables2clouds.com
Follow us on YouTube: https://www.youtube.com/@cables2clouds/
Follow us on TikTok: https://www.tiktok.com/@cables2clouds
Merch Store: https://store.cables2clouds.com/
Join the Discord Study group: https://artofneteng.com/iaatj
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Tim (00:13):
Hello and welcome back to another episode of the Cables to
Clouds podcast.
And uh with me as always is myco-host, Chris Miles.
I am Tim McConaughey, and wehave a guest with us uh from
Learning at Cisco, uh JasonBelk.
Jason, go ahead and introduceyourself.
I don't think you've been onthe podcast before.
Jason Belk (00:31):
Hi.
Thanks, Tim.
Thanks, thanks, Chris.
Yeah, I am Jason Belk.
I am a senior technicaladvocate with Learned with
Cisco.
Um, the people who do thetraining and certification, CCNA
up to CCIE and everything inbetween.
And uh a little bit aboutmyself.
I live in Reno, Nevada, married12 years, have two young
daughters, a black lab, and uhmy background is in I've been at
(00:52):
Cisco since 2010, off and on.
And I I started out at Cisco ITas an analyst, network
engineer, network automation,started teaching networking
automation before DevNet to myour networking team, learning it
myself, and then I left andwent to network to code for a
few years.
Um, came back to Cisco as a indeveloper relations in the Cisco
NSO BU, which is like basicallya big automation product, and
(01:13):
then now more recently inlearning with Cisco.
So kind of done around thecircuit inside and outside of
Cisco and have some operationalchops to show for it.
So I kind of know the strugglesof all the operators out there,
uh at least for the on-premisestuff and the enterprise.
I was on the campus and branchteam when I was on call.
Um, so yeah, and and I'm reallythrilled to be here.
I mean, we've had conversationswith with you guys off and on
for a while now, and big fan ofthe show and love what you guys
(01:36):
do both for the news perspectiveas well as covering, you know,
all the different stuff fornetwork engineers who are trying
to grapple with AI and cloudand everything else.
Thanks, man.
Tim (01:45):
Um, so yeah, yeah, what's that?
Oh, I just said I appreciatethat.
Oh, appreciate that.
Yes, we are appreciative.
Uh it's always nice to meetsomeone who actually listens to
the show.
I'm just kidding.
Um anyway, so we broughtactually we brought Jason on uh
because he is working onsomething that's actually pretty
(02:05):
cool.
Uh we've been we've beentalking it back and forth for a
long time, actually.
We've been having trying to getuh Jason on the show, so we
finally made it work.
Um and Jason is here to talkabout something that I didn't
know existed, and I guess thatprobably a lot of the listeners
were not aware existed.
Uh something called spec drivendevelopment.
And uh I believe it hassomething to do with with
building for AI, but actually,Jason, I'll you're the expert.
(02:26):
I'll let you I'll let you takeit from here.
What is spec drivendevelopment?
Jason Belk (02:29):
Sure.
I mean, I've been using it forabout six weeks.
Uh and AI speak that makes mean expert.
Um that's right.
Uh to be fair, no, like one Iactually asked Claude Coder or
today how much I'd been using itjust on one project, and it
said 150 hours in the past threeweeks.
Um project, one one personalproject.
(02:51):
Um and I've and I I have likefive or six personal projects
going on top of some work stuff.
And and so all that to say iswhat is spectrum development?
I mean, if you think if you'vebeen in IT or any amount of
time, you're familiar withproject managers.
They're those friendly peoplewho help keep you on track and
get all your T's crossed andyour I's dotted.
And if you if you've tried tobuild anything of significance,
(03:13):
you know it helps to have aplan.
And so when you're talkingabout an AI, you you I think
we've all heard of the exampleof AI being, you know, like an
expert idiot or something.
Like they have tons ofknowledge, tons of ability to
execute, but what I would say isvery poor wisdom and judgment.
And so how do how do you helpsomebody who is able to do
whatever you say and sometimesinterprets it a little too in a
(03:37):
direction you weren't expecting?
Is that that you give them aplan.
And there's definitely thisisn't the only way of
approaching that problem.
And this isn't for um, youknow, I'm on chat GBT, Jim and
I, Claude, you know, I'm justdoing my chat bot and I'm
talking back and forth.
I'm not talking about thatscenario.
The the way that I'm I'mtalking about specifically is is
for a quote unquote agenticcoding, um, where where you are
(03:58):
giving prompts, instructions toyour computer through a text
interface, in my case, usually aterminal, and you're you're
typing in just as you would onChatGBT or Claude or Gemini, and
instead of getting a responseback that says, hey, I can't do
that, sorry, it actually goesoff and tries to do it using
your computer resources.
Um so you have to havepermissions and trust in the
(04:20):
place where it's running for itto execute that code.
And by default, something likeClaude Code, which is one I'll
be referencing in the rest ofthis conversation just for sake
of simplicity, um, will ask youpermissions when it wants to
create files, when it wants todo web searches, when it wants
to edit edit or delete files, dobash commands, stuff like that.
But all that to say is that itcan do all those things for you.
(04:40):
And not just from a codingperspective.
Um, there's another podcast Ireally enjoy called Software
Defined Talk.
And they had a whole episodeabout cloud code I'd recommend
people check out.
If you haven't heard about itbefore, played around with it
before, it gives you a goodprimer of what I would say
somebody who's not a coder or anactive coder using cloud code
just for productivity stuff,like sorting your screenshots in
your screenshots folder andnaming them, um, or uh, you
(05:02):
know, putting together yournotes and things like that.
Because it can basically onceyou give it access to those
files, and that's of course atrust issue, it will then be
able to do instructions based onrenaming, reading, and all that
kind of stuff.
So going back to spec drivendevelopment is that okay, you
you give this a local AI agentthe capability to do things on
your behalf.
And if you start asking to dothings, as I think we've people
(05:24):
who have worked with AI for awhile will experience there's a
context window issue.
And I'd say there's also like ascope issue.
And like I meant to do this,but actually you started doing
this other thing, and thenthat's not too bad, but actually
you were kind of deviating fromwhat I originally talked about,
or you built a production gradesystem, and I was just trying
to figure out if it worked.
Um, and or most common one Istill even run into, even with
(05:46):
spec driven developmentsometimes, is that you're trying
to pip install this on mysystem Python, and I want to use
UV.
Um, so spec driven development,not to say it solves every
possible problem, but I've gonefrom like having to intervene
and course correct my quad codemultiple times per session to
like maybe once a week.
And I can just let it run foranywhere between 30 seconds and
(06:08):
a minute.
I've had it sometimes runningup to several hours.
Chris (06:12):
So it's funny you mentioned that episode of
Software Defined Talk.
I just listened to that lastweek and it was uh it's it's
it's funny because like theythat was like my motivator was
like, man, I I really need todig in with clawed code because
they even call it out in theepisode, but like the fact that
it's called clawed code kind oflimits what you think in your
brain that you can do with itbecause you can do very vast
(06:34):
things with it.
But um, yeah, I like I likewhat you said about the the idea
of you know spec drivendevelopment kind of being that
specification that you define upfront to kind of keep your AI
in line.
But I'm I'm curious, like uh Ifeel like I've I've heard of
spec-driven development uh uhyou know several times,
obviously very recently as well.
Um, was this the thing thatjust came along with AI almost
(06:58):
as a not guardrails, but kind oflike guiding light for AI um
coding?
Or was this there before?
Jason Belk (07:05):
I'd say that the principles, um, and and I'm I'm
we'll share some links in theshow notes that people can take
a look at.
Um the principles I think goback to project management
principles in terms of definingrequirements and and and
executing executing upon them,uh, you know, clarifying
assumptions uh in a more formalway.
Uh, but I'd say, at least frommy knowledge of it, I've been
following it since August of2025.
(07:27):
I think that's around the timeit came out.
Um, because I I follow a lot ofdifferent YouTube channels.
One of them is like the GitHub,you know, AI folk, and and they
had some Microsoft DeveloperDay back then where they were
walking through it.
Um and so the the the one thatI use is the GitHub, what they
call spec kit, and it's onGitHub for free.
(07:47):
I'm I'm also sharing, um, I'llshare a link to my personal
spec-driven development kind offork of it, um, with the caveat
that the cut the code I have inthere that is the spec driven
development part is uh is notlike updating with the remote
fork.
I haven't got that complicatedon it yet.
So if you want to have thelatest GitHub spec kit part, you
might have to swap that partout and then use the rest of my
(08:08):
bits I have in there.
All that to say is um it's it'sstill relatively recent, as of
August.
And then I I tried to get itset up myself a couple times,
but I'd say the documentationout there for me as an
intermediate level programmer,novice level AI person who is
figuring stuff out as long as Ifor the past since 2023 or
whatever, since GPT 3.5.
(08:29):
Um I I I had a couple of stopsand starts until this guy named
Jason Gekki, who who's um I I'mconnected with who's people at
work pointed me to some umdocumentation that he does.
He basically he created apresentation that is happy to
share with whoever it wouldhelp.
And he's using this a lot, hisstartup and and consulting.
And so I'm gonna share the linkwith that as well, which I
think just helps visualize theflow and how all the pieces fit
(08:52):
together.
Because if you just go to theGitHub repo, and this is where I
got stuck, I'm like, okay, sowhat do I do with this?
I I see we have a repo and Isee I have clawed code, but
like, where do the dots connect?
Chris (09:01):
Yeah, for sure.
Jason Belk (09:02):
And and the first couple of times I tried, I kind
of gave up before I got throughhis stuff.
Um, not through it in a badway, but like before I I I went
through his his the way heexplained it, and then I've
tried to explain it as well onCisco U and a tutorial we'll
talk we can talk about a littlebit later, is just kind of like
connecting that dots of like,okay, how do I actually use
this?
And like it's not it's reallynot as mystical as it sounds.
It's and once you I I even wentthrough a demo of this with one
(09:26):
of my friends from church, andhe's he's like he has some ideas
on on a potential businessidea, and so I I shared my
screen and my laptop on thecomputer, and I literally did
the spec spec flow with him forhis idea.
He was defining therequirements verbally to me.
I would just type in whateverhe said, and then we literally
built a product in like 45minutes.
Tim (09:41):
So help me out here because um I'm new to the spec driven
thing.
Sure.
Uh the spectrum, so you know,I've been messing around with
cloud code, I've done somethings, I've failed at more
things probably than I'vesucceeded at.
Uh I tried to make a news agentfor the show and I posted that
about that on Substack.
Um if you know, me beingknowing as little as I do, but
(10:03):
and and knowing, you know, beingdangerous with it, I
supposedly.
How do I how would I, as aperson who's messing with it,
getting better but not great, umintegrate?
Or where do I even start tryingto pull in spec driven
development into like my ownworkflow to build stuff?
Jason Belk (10:19):
Sure.
I I would say start with agreenfield project, something
that like you had an idea.
It's you're not trying to boilthe ocean, like I need I need to
make the next Facebook app orsomething.
But if you're an infrastructureengineer, you're like, okay,
maybe I I'm not supercomfortable with Ansible and I
want to deploy some Terraform aswell.
And like, okay, uh, or I knowTerraform really well and I
don't know Ansible.
And so it's like you have anyou have a context area that you
(10:40):
already kind of understand, butit's kind of rusty and you have
to look up a lot of docs.
So you're able to define therequirements reasonably well.
So I'd say, like, likebrainstorm, okay, this is what I
kind of want to do.
And you don't have to get itall right in the first shot.
This is more just kind of liketo make sure set you up to
succeed.
Now I'd say check out the linksthat we were we set up because
it's it's something that I'lltalk through, but it's almost
more important for you to lookthrough and do it yourself with
(11:01):
your fingers on the keyboard.
Um, because it it's it's reallynot rocket science.
Like once you have on onescreen the flow and the other
screen your terminal, and likeyou have it set up.
Okay.
The the installation process isbasically just a bunch of bash
scripts.
It's not like installing anynew software, it can kind of run
independently too, just as it'sa bunch of markdown files and
bash scripts that kind of likeguide the quad code to do the
(11:23):
the project management workflow.
And so the the the GitHub repothat I have for for my um kind
of additional context, I'll justsay some additional
instructions on the spec kitflow, which is optional, or the
core one that you can get fromthe official Microsoft GitHub.
Um you can basically get clonedown the spec kit.
I'd say the one thing I learnedthe first time I did that I did
it incorrectly is that Iinitialized my clawed code at
(11:44):
the wrong directory because Iwas kind of new to Claude Code,
where I was like, okay, uh I'm Ihave my project folder, make
dira project, and then I putspec kit in that in that
subfolder, and then I ran clawedcode at that layer above where
the spec kit was installed.
So it didn't look because theseare technically what are called
claud skills, and it doesn't itdoesn't load in the skill if
you're not in that directorywhen you initialize claude code,
(12:05):
which it didn't tell me.
And I mean like I I went allthat to say is like if you're
doing it, just make sureinitialized when you type in
claude on your terminal when youhave it installed, do it in the
same place, the same layer ofthe directory where you have the
spec kit installed so it canload in those skills.
Tim (12:19):
Oh, okay, I see.
So so I mean you're pullingdown a repo, like you said,
basically, and and and in thatrepo is gonna be like skill
file, like files essentiallythat that define all the skills
for that you're gonna use forthe for the workflow that you're
gonna follow on the spec drivendevelopment.
Okay, that that makes it alittle bit more real to
visualize.
Okay, I see.
Jason Belk (12:38):
And you can inspect them, and it's not like some
binary that you can't evertouch.
You can look at the bashscripts, you can look at the the
markdown files that are a partof it.
Um but basically as soon as youload up that clawed instance
and it loads in those skills,you can then access on slash
commands so you to to load it touse those skills.
So you go, you know, like slashspec kit dot, and then the
first one would be constitution.
(12:59):
And if you do it in the wrongorder, it'll kind of like gently
correct you.
It's like, oh, you're trying toimplement, you don't have a
plan yet.
Tim (13:05):
You want to have a plan it's a workflow, it's yeah, it's
a whole it's like uh uhwaterfall.
Yeah, yeah.
Okay, yeah, that makes sense.
Jason Belk (13:13):
Yep.
So so the the the first step,and this is one that I'd say it
doesn't always enforce, but it'sgood to start with because it
kind of gets you in the rightheadspace and helps it up
farther down the road, is theconstitution.
So just like in the UnitedStates and other countries, they
have a governing body thatdefines the principles of what
should be executed.
Uh so in in this case, it'sbetween I think three and five.
(13:34):
Generally, it suggests.
And and it's not that you haveto walk in with like a quill and
like some parchment and and andand have it all worked out
before you start.
Like Hancock on there.
Chris (13:44):
Yeah, get your drawing Hancock.
Jason Belk (13:46):
That's right.
I actually use um this is kindof like a JSON hack that I
started recently, one of myfriends pointed me to as Whisper
Flow.
This is different than spec hitdevelopment, WISPR flow.
It allows you to do voice totext real easily.
So you have to have a hotkey,and you can do that prompting
into your terminal.
So it makes my prompts so muchbetter on Claud Code, where I'm
not like getting carpal tunnelDM VPN on my on my Claud Code.
(14:11):
I just hold down option, that'sthe key I set up, and I just
talk for like 30 or 45 seconds,brainstorming what I want to do
at the consultation at theconstitution where it'd be more
at first principles like, okay,yeah, I'm trying to do a in this
case Ansible Terraform app.
I kind of I know Ansible reallywell, but I don't know
Terraform so well.
Um my my resources are allgonna be public cloud, AWS,
Azure, and like my keys aregonna be in this.m file or in my
(14:34):
you know my my ZH S HRC and allthat kind of stuff.
So you just kind of like giveit the lay of the land and kind
of what direct what whatdirection you're heading.
And I just kind of record 34 to5 seconds, paste it in there,
and I'm done.
And and then it processesthrough there and uses the the
LLM.
And that's a one other keything for me is not to use
sonnet, but to use Opus.
(14:54):
When I did spec drivendevelopment, like it default.
I mean, I think quad codedefaults to Sonnet, which is
fine for like you know settingsimple stuff, but you can do
slash model and switch to opus.
And yeah, you're gonna use upyour tokens a lot faster, but
it's like cry once, try once, orwhatever, and like because
otherwise it was just spinningin circles with Sonnet.
It's like, oh sorry, I brokesomething.
Let me fix it.
(15:15):
Oh, I broke something else.
Let me fix it.
And so I was like, with Opus,it still breaks things on
occasion, like once a day forme, or and what stuff like that.
But I'd just say, even if youhave less Opus tokens based on
your pro or whatever level, justwait till it refreshes.
Yeah.
Tim (15:30):
Yeah.
So this is interesting becausewhen I was working on my agent,
um, I was trying to save savetokens.
So like I was using um I usedhaiku, which I was like, oh wow,
this is a really bad idea.
It's a really bad idea.
I realized it after the fact,after I was building the code
for it, I was like, wait aminute, I shouldn't use haiku to
build this.
But what haiku is good at iswriting.
(15:50):
So maybe from like this part ofit, this constitution defining
like doing the writing part ofit, would that be better just to
get it all in the mod?
Like get it because all you'redoing at this point with spec I
think, and correct me, is atthis point, really what you're
doing is creating the files thatclause refer to uh refer to for
(16:11):
the development part of it,right?
Jason Belk (16:12):
Yeah, yeah, I'd say I haven't tried that.
It's a good idea.
Like in terms of the flow, I'dsay definitely when you get to
the implementation layer, youwant to switch to over.
But you could probably get awaywith it on Sonic and Haiku on
the on the initial ones.
I I haven't done like A-Btesting to see what the
difference is, but I think formost, I think on the
implementation level, that'sdefinitely for your own.
Tim (16:29):
Oh, it's definitely no doubt, no doubt.
I can tell you haiku isterrible at implementation, not
accident.
I can tell you that for sure.
Jason Belk (16:35):
That's a good idea.
Um, and so yeah, for once youget your constitution.
So I I do my brain dump, itresponds back basically, usually
like, hey, I've noticed basedon what you said, or if you
didn't give enough information,I'll ask them follow up
questions.
Um, based on what you said,here's three to five principles
that I'm seeing.
Like that you you want to havesimplicity as a design, you want
(16:56):
to have cloud native as adesign, you want to have these
you're doing this as a proof ofconcept rather than a production
build.
So like it kind of extracts outor clarifies for you what type
of project you're trying tobuild.
And this doesn't just have tobe for coding.
Um, this could be for for otherstuff too.
Like that's but all all that tosay is like what then it
creates from the that response,then it generates your
(17:16):
constitution file for you.
So I'm not like going in therehandcrafting.
You can you can go in there andedit the file, but it basically
writes it for you.
And then you go to specify.
So once you have yourconstitution going, you flow
through through these this loopof this workflow for each of
your features.
So you go from specify to planto tasks to implement, back to
specify to plan to tasks toimplement for each of your
(17:39):
features, basically.
Um, and and and I've found forbigger projects, oftentimes my
first spec is like anarchitectural one where its
output is not code, but justlike architectural documents,
um, where I'm like, okay, let'sspecify what cloud we're gonna
be using, what you know, whatlogging we're gonna be using,
what um all this kind of stuff.
Where uh if it's a much bigger,if you know it's gonna be a big
(17:59):
project, that'd be myrecommendation because sometimes
it just it just makes decisionsand does even with the spec kit
flow is gonna start making someassumptions on things that you
might have to go back and reworklater, which isn't in the
world, but like that'd be my onerecommendation for a bigger
project.
But if you're just gettingstarted doing a small thing like
this, it doesn't matter.
I would say you're probablyjust gonna have maybe two or
three specs tops.
Um so the first the first onewould be, you know, like the
first one that flows.
So we have the constitution,and then you go to and then you
(18:21):
do the slash command slash speckit, and it'll have like tab
complete if that skill is youknow um loaded in there
correctly.
And then it sometimes reallyit'll even suggest, like at the
end of its constitution, like,okay, we have the constitution
already.
Do you are you ready for speckit specify and like copy and
paste?
Yep.
Um and and and so then for speckit specify, you're documenting
your requirements.
So going back to that projectmanagement example, you're
(18:43):
sitting down with the projectmanager, and they're like, Okay,
so what's what's the problem?
Like, well, we have a new salesoffice and we need to get
access in for salesforce.com.
And they we we have an appwe'll have to take down the
network over the weekend inorder for this this network
change to take place.
Um, and no, we're gonna do iton the Cisco iOS cat9 gates and
(19:04):
Cisco Catalyst Center orsomething.
So you're just defining youryour requirements in terms of
what you're trying to accomplishfor this particular feature.
Tim (19:12):
Yeah, that makes sense.
I mean, you're setting up theguard essentially the
guardrails, like what are yougoing to be working with?
What are you gonna be doing,right?
Chris (19:19):
Like I think uh I think one thing that I've definitely
observed is it's amazing to mehow fast Claude and Anthropic
have kind of switched to thefront runner in this in this
conversation.
Um it's it's been you knowspecifically over the last it
seems like over the last monthlike it's really really picked
(19:39):
up steam.
Um but one thing I've alsonoticed is the the kind of you
know kind of variety of toolsthat people are using.
Like you said, you're usingsomething specifically for voice
to text, and there'sintegrations and there's
people's building skills andusing different MCPs and things
like that.
So maybe can we can we take astep back and let's talk a
little bit about skills andbecause I know with with spec
(20:02):
driven development, um it seemsone of the kind of concrete
things is is the idea of becauseyou've well defined your
requirements and your intent upfront, this allows for kind of
parallel development, right?
You have this you have agentsworking on one piece of it and
agents working on another pieceof it, and they can operate
autonomously, right?
(20:23):
And they won't conflict witheach other because they all are
working off of the samerequirement, right?
Correct.
Um, so can you tell me a littlebit how agents and skills play
into that?
Jason Belk (20:32):
Yeah, I I I'm by no means an expert on um skills,
but I'll my humble explanationis that they're just a series of
bash scripts and markdownfiles, primarily markdown files,
that um help the model, usuallyClaude, understand what parts
of its training and brain to usein that particular context.
So just like if you're onChatGBT and you say, pretend
(20:53):
you're an expert networkengineer and analyze this ACL.
Okay.
Um it's it's it's it's it'sit's it's funny because like on
the software-defined talkpodcast, they actually even
talked about like if you if youlook at like the market share
that's been destroyed by a bunchof markdown files, they're
like, it's tremendous.
They're like, you know, you'redestroying accounting, it's
destroying legal.
It's like you look at whatactually is destroying, it's
just it's just like somemarkdown files on these repos.
(21:15):
It's not like they they didanything super special, it's
just like it it's such apowerful model, it just needs
like little guardrails and likesome to say, okay, just focus on
this part and take on thisidentity.
Um yeah.
Tim (21:29):
So actually, real quick, because uh you just reminded me
of something, and I I this ismore of an opinion thing because
like I don't think anybodyknows what's gonna happen in the
next six months or three monthsor two months with AI, these
things.
But isn't this this feels likeuh, you know, what I'm hearing
more and what I truly believeprobably will be the the result,
the end result here is as weseek to make these things more
(21:49):
efficient, and as trainingbecomes things that, you know,
we have enough material uhhardware to start doing more
training, specialized training.
Do you think that we're gonnagravitate towards more like
small, you know.
Small specialized models thatare going to be built kind of
for this, or unless, or do youthink it's going to be like, oh,
well, here's an agent, we'regoing to put the guardrails in
place, and then we're just goingto have the LOM still back at
(22:10):
all?
Jason Belk (22:11):
I think it's going to, my opinion, just the way
I've seen other stuff work,you're not going to know or care
as a user.
Sure, no, 100%.
But but but like from atechnical perspective, um Yeah,
that's what I'm thinking.
I I I I can definitely see thatbeing in the approach.
I mean, that's why I work atCisco, like they released a
security um L um special trainedmodel.
Yeah, a small language modelfor that last June on Hugging
(22:34):
Face.
Because uh yeah, like I thinkit's also just not feasible to
maintain these models.
I mean, I might be wrong, I'mnot an expert by any means, to
like try to pretend like thecorpus of every possible job in
existence that will exist andthat will exist, could be
maintained by something that orwould want to be maintained,
even if we could, it's like,well, we could we could use all
(22:55):
of the power from Idaho to likepower this thing, or or or we
could just run it on your run onyour phone.
Tim (22:59):
And so that's kind of my feeling too.
Like, so I I think that'sprobably where we're going to be
headed once once we get more ofthe of the process of being
able to truly kick crank outthese small language models or
specialized language models,depending on how you want to
talk about it.
Jason Belk (23:14):
Yeah, I think it's right now such a new thing for
so many people and so disruptiveto people's workflows and such
an open form thing.
Or previously, with like cloud,you're like, okay, I have to
get like my AWS keys and likeyou know, go onto the console
and like figure out what I wantto do here, or or even from a
SaaS perspective, like okay, Ineed to like get my credit card
and like get this SaaS thingplugged into my work workflow.
(23:35):
And that there was still likesome complexity there and like
some changes to people'sexisting paradigm.
They're like on-prem versuscloud, hybrid, you know, all
that kind of stuff versus likeLMs just blows everything out of
the water.
I was like, what if you couldjust like talk to your computer
like in Star Trek and it couldlike answer?
Right.
And it's just like I I I Imean, I'm not trying to like
trivialize anything about this,but I think that's in my mind
(23:57):
one of the foundationaldifferences that it's so open
form and free form that it'shard to train people on like
what to do.
You're like, oh, I you mean Ican have it help me cook dinner
or build out ACLs on my router.
Okay.
Um, but I guess we'll do that.
Um, so I I I think that's mymind what the adoption curve is
(24:18):
gonna is is is gonna eventuallylook like is that it's gonna get
more adopted as you get morespecialized bottles where people
can understand their use case.
They're like, oh, this is mysecurity model, I'll ask it
security questions, even thoughthey could probably do that
right now.
Tim (24:29):
Um yeah, yeah, yeah.
But efficiency-wise, of course,you gotta when you every time
the ask and LM a question,you're lighting up like half the
planet to to get the answer.
Um you said something earlieralso about the um oh about the
uh oh man, I just lost my trainof thought, of course.
Darn it, I I had it for asecond and then I just lost it.
(24:50):
I give up.
Uh Chris, uh I'll head it overto you.
I was I was kidding.
Jason Belk (24:54):
I was I was put I was put in the spot for no
reason for that.
Tim (24:57):
Sorry.
Jason Belk (24:58):
I I can keep going through this stuff that's uh so
so we we we coveredconstitution, we covered
specify.
So after specify, you have theplanning phase.
So you've defined yourrequirements, you've defined
your governing principles, andthen now you have your
requirements, you go per featuredesigning your technical
approach.
And and and this would be justthe next flash command.
So after it does the specify,and that might take several
(25:20):
minutes or longer in terms ofthe specifying the constitution.
So you gotta be patient, go tothe bathroom, take the dog for a
walk, um, and then come backand then do the spec kit plan.
And you can do these inmultiple tabs.
Like I'll sometimes have threeor five different spec kits
running on different projects atthe same time, which just makes
you feel like superhuman.
Um and so, like spec kit planper feature design technical
(25:42):
approach, and that that'sbasically like, okay, now that
we have our requirements, likewhat's gonna be our approach for
the Ansible Terraform example?
It's like, okay, you know, likewhat modules are we gonna use
or for Ansible and and what umproviders are we gonna use for
Terraform?
And this isn't something thatlike you have to define that out
because like it based on theLM's knowledge, and if it has
access to the internet, it canfigure that stuff out.
(26:02):
So, like this plan isgenerally, in my experience,
generated from the your maininput is on the specify, and
then the plan it builds a planfor you.
You're not like okay, I have tospend half an hour figuring out
my plan.
It's like, hey, based on yourrequirements, this is why I
think a good plan would be basedon your governing principles
from your constitution.
And you're like, that is a goodplan.
Let's do that one, or you cango in there and you can clarify,
(26:23):
and this is for the voice totext input, I'd be like, I'd be
like, no, actually, like Idecided not to use Ansible.
We're we're we're gonna usePuppet and Chef instead.
And it'd be like, okay, I guesswe need to go, we we need to
rechange everything.
And so, like, you you can stillpart way through change your
mind and do different things,but or maybe a more realistic
change to the plan would belike, okay, we're gonna have
this on specific jump posts thatyou need to know about and and
(26:44):
and and stuff like that, havingmore details on your on on what
is gonna affect the plan.
Tim (26:49):
Actually, you just reminded me, just reminding me what I
lost my frame of thought on,which is what we were talking
about, and and I think everybodywho's probably now put any
amount of time into AI isfiguring has already figured
this out, but just for settingpurposes, um, the more you know
about the technology you want touse, the easier it is to get
the AI to to under because ofcourse, because you can say,
(27:12):
like, oh, I know you're gonnaneed this information, right?
I know you're gonna need toknow about this, I know you're
gonna end up like if you knowAnsible, like if you know enough
about Ansible, you know exactlywhat that you gotta tell make
makes it into the spec.
Yeah, you you know exactly whatneeds to go in there.
Jason Belk (27:26):
Yeah, yeah, you know what the gotchas are.
And so this is where I thinkpeople who don't have a
technical background use AI, dokind of get caught up or get
stuck or get in rabbit holes.
And so this is like exactly toyour point, like your advantage
as a technologist in definingthese things is that you can
front load your assumptions onall the things you wish, as if
you were talking to yourselffrom a mentor version of
(27:48):
yourself where you've forgotteneverything, but you also you're
like this is yeah, or a junior,a junior engineer, and so you
need to clarify yourassumptions.
And and this is kind of mything that I've realized over
the past year, year and a halfis that like your leadership
skills, your soft skills as ahuman being, make your ability
to execute prompting andplanning like this uh
significantly better.
(28:09):
So I'd say brush up on somebooks on leadership if you have
nothing to do and you're boredon the technology side, and it's
gonna help you be a better AIuser in the future because it
gives you things like clarifyingassumptions, you know, defining
requirements and all thesethings that like project
managers and people with withyou know technical leadership
(28:30):
backgrounds are all used to orconsultants oftentimes too,
where you have to sit down anddefine your workflows.
Um it's uh it's all in yourhead somewhere where you you
need to get it on paper, kind ofassuming nothing.
Chris (28:40):
Yeah, I think um one thing I've definitely noticed in
working with AI forspecifically this type of uh
this type of use case is uh I'mvery good at giving context, and
I often give a very largeamount of context, which as you
guys talked about, it really uhit really messes with your you
(29:00):
know your windowing uhavailability when you're
interacting with AI, so it'slike burnt me a few times
because I'm like I'm like no,this is important, you need to
know this.
So I do like this idea.
Yeah, I do like this idea ofbeing able to define that up
front and that just using thatas a reference source over and
over again.
Um because yeah, I mean, butthen there's tools as well out
(29:21):
there, like it we should weshould put a link to that
software defined talk episode inthe show notes because it's so
valuable to me.
Um but there's um theyreferenced that there's even
tools out there that you can usethat actually remove all the
fluff from any of your promptsthat like can't will not, you
know, so that you're not eatingup tokens that's cost.
(29:42):
You might be saving them.
Yeah, exactly.
Like you might be using themuh, you know, maybe 10% less,
but that 10% can be you knowthat's valuable for you down the
down the road, right?
Jason Belk (29:52):
And I'd say if you have other documents, you can
reference those as long asyou're you're willing to give
access to those files to thissystem.
Right.
So like for one of my projects,I had done some work in a
planning phase, not even beforeI even knew about Spectre
Development.
I I had like a 20-pageMicrosoft Word document that I I
had defined basicallyrequirements.
It was the first half of it wasgoverning principles, and the
(30:12):
second half of it was executionlines, not like code, but more
like defining tasks and things.
Yeah.
Um, and I p I found PDFs to bethe longer franca of of AI, and
so I exported the doc.
I think you can read doc X, butI always just do it as a PDF
just to make sure theformatting's right.
So I export as a PDF, includedthat in my constitution
specifying that like it wasreally good at that, because
(30:34):
it's like it it extracted allthe important details and kind
of like filled out the stuff forme on that stuff.
So if you already have aproject plan from like an actual
project manager, and they'relike, okay, okay, Chris, we need
we need you to to build outthis cloud deployment.
Here here's your here's yourproject plan.
You're like, okay, I'll exportas a PDF, and if my company lets
me use these types of tools, Icould just use spectrum
development to extract out andbuild out the plan from their
(30:57):
plan.
Tim (30:58):
Yeah, that makes sense.
And I don't run into that.
Jason Belk (31:01):
Oh, go ahead.
I was just saying once so onceyou one thing is that it
sometimes your plan might haveambiguities where e either um
based on what you're trying toimplement or the different
technologies and outcomes you'retrying to achieve.
And I'm trying to think of agood one.
Um, but an example might be ifyou weren't specific on a cloud.
If you're like, okay, I want todo Ansible and Terraform, but
you never really said if it wasAWS, Azure, or something else.
(31:24):
And it will oftentimesauto-suggest in at the end of
its response, like, do you wantto go to the next phase or do
you want to do spec kit forclarify?
And you can still run it, evenif it doesn't suggest it, and it
might say, Nope, there'snothing just to clarify, but uh
sometimes it will it willactually go through a clarify
phase where it goes through allthe tasks, I mean all the plan
that it's built and say, Here'smy plan.
(31:44):
There are actually someambiguities on some of these
plans, and then it will it'll dobetween one and five questions
that are multiple choice, or youcan fill in your own answer and
and it and and allows you toclarify.
So it's like like, for example,those questions you'll say,
like, I noticed you know ourplan we're we're we're we're not
saying which cloud we're using.
Do you and it'll kind of knowbecause it's smart in terms of
(32:06):
like, did you mean to use AWS A,B, Azure, C, you know, GCP, or
D, something else?
Fill it in.
And it doesn't have to be aone-word answer.
You could I can say, well, I II I have an Azure account, but I
prefer AWS, and so it'll belike, okay, no, let me update
the plan.
Tim (32:20):
Yeah.
I like to use the uh I mean, Ithis is uh obviously spec driven
is much more uh not justcomplicated, not just complex,
but obviously very, I assume inthe weeds with it.
But I mean I've been using theplanning mode of Claude quite a
bit.
Um for the it for the same forthe same reason, right?
Create me a file because I knowwhat's gonna happen is that I'm
(32:42):
gonna run out of context whenyou can compact the
conversation.
I'm gonna lose all the detailsI put in.
Exactly.
Go back and refer to this fileagain when when you've when you
after you've done that, right?
Jason Belk (32:52):
Exactly.
And that's one thing Imentioned early on, but I forgot
to mention again, is that thatis the huge advantage of this or
the planning mode on ClaudeCode is that once you get to a
certain point, all its tokens inits brain get scrunched down to
a little dot, and and then ityou you kind of have to start
over again to certain steps.
But when you have this plan allwritten out, it just refers
back to the files.
It kind of knows where it is inthe flow, it has like its own
checklist.
Like, okay, I'm on the planningphase.
(33:12):
Now let's move on to the tasksphase.
Tim (33:14):
Yeah, it's a it's a big big bonus.
Um I I yeah, I had to learnabout it uh because I didn't
realize that until I did threeor four projects and I was
losing all my context and it wasforgetting everything.
And I was like, man, there'sgonna be a better way.
And so yeah, that's the anyway.
No, this is good.
Um we've got all the so we haveall of the the links, including
(33:35):
the one for uh that goesthrough the entire process of
spectrum development.
I think that's gonna be use uhreally useful.
I think we just that I thinkthat's probably the first thing
where people what I'm thinkingabout is okay, there's two
things, right?
One is where do people start?
We kind of covered that, whichis like literally just go look
at these documents and then gofork a repo or grab download the
repo.
That's so that's just to thelisteners that are already work
(33:57):
with AI or that want to startworking with AI and want more of
a framework to start from.
That's the first one.
Um, the second one is uh sincewe have to we'll we'll start uh
wrapping up here in a minute,but let's talk a little bit just
before we wrap about the CiscoU cores that you built.
What's at like compared to theto the docs that are the the
(34:17):
docs we were like to see?
What is the what covers what'scovered in the course?
Jason Belk (34:20):
So so yeah, I we have a couple links here.
We have one that's a tutorial,these are free.
You don't need anysubscription, you just need a
CCO Cisco account, which doesn'teven have to be a Cisco
customer.
You just you can create one forfree with your Gmail, I think,
or whatever.
And and so this to Cisco Ututorial, Cisco U is our
learning and certification,learning with Cisco, place where
you get your certificationtraining officially from Cisco.
(34:40):
So they have you know CCNA upto I think CCNP and and other
training in between.
Um and so we also have freetutorials on a variety of topics
from cloud, security, AI,networking, the gamut.
Uh I think we have over 200 atthis point.
So not just this one tutorial,but check out the other ones.
We have ones linked todifferent certifications you
might be pursuing as well.
So for for this particular oneI wrote a few weeks ago, I
(35:03):
basically took the spec drivendevelopment um gecky link that I
share, I talked about earlierand kind of converted that to a
network engineering context.
So if you uh if you look at hisgreat presentation link and
you're like, okay, I kind of getwhat it is, but how do I do
that as a network engineer?
Look at my link on the Cisco U1where I basically have you go
through a spec driven flow forsetting up the skill and then
(35:23):
using it to like basically buildlike a simple Ginja 2 YAML
parser thing that creates someconfig files.
Um, so that should get you upand running, at least with a
very simple use case that's allgonna be run locally.
You don't need any remotedevices, you can just kind of do
it with your Claud code locallyon your personal laptop or
whatever to get started.
Uh and then I'd say on top ofthat, so that's the free one on
the paid side, and I think theymight have it free right now,
depending on when this getspublished, of a Cisco AI
(35:45):
technical practitioner course.
And we came out with acertification last December.
So, this in terms of complexitylevel, on those who are
familiar with Ciscocertifications, the CCNA Cisco
certified network associate isan associate level exam.
This is one step before that.
So this is meant to be entrylevel.
I don't know what I'm doingquite yet.
Let me figure it out, type ofcourse and exam.
So we came out with a courseand exam in December.
(36:06):
So it's been out for a fewmonths now, and it it covers
prompt engineering, you know, AIsecurity, AI-driven software
engineering, um, advancedprompting strategies, uh, you
know, what is an LLM, what isRAG, and all that kind of stuff.
And I'm I'm actually doing aCisco Live session covering this
AI tech course andcertification.
So if you're coming to CiscoLive US, US, come check out my
(36:27):
breakout session with withGeorge.
Nice an hour, it'd be a lot offun.
Um and then I'm also doinganother um Cisco Live session on
the new um CCMP automation,autocore, EN core, like the
enterprise networkingautomation, professional level
that used to be called DevNetProfessional Exams.
So I have another one therethat people should check out.
We're running Cisco Live, andthen third, I have what's called
(36:48):
Cisco U session.
So we have 45-minute sessionson on the floor, and that one
will be on Spec Drivendevelopment specifically.
So if you're still unsure andyou happen to be a Cisco Live,
check out my session there orthe recording will be up, you
know, after Cisco Live as well.
Awesome.
Chris (37:01):
Yeah, we'll definitely get all those shows or all those
links in the show notes aswell.
I think this will definitelypublish before Cisco Live.
So yeah, for sure.
If you are if you are going toCisco Live US, make sure you do
check out the session.
That sounds great.
Is it our details?
Do you have details of thesession as well?
We can put the I can get this,yeah.
Yeah, let me pop.
I can pop the side.
Tim (37:20):
Yeah, the catalog is up, catalogs up now, so we should be
able to we'll we'll get that.
Um I'm doing uh I'm doing alab, but I'm not here to plug.
I'll plug my stuff on adifferent show.
But uh yeah, so we'll get we'llget that.
We'll get that in the shownotes.
Uh definitely sign up for allof Jason's stuff.
Um also, Jason, where canpeople find you online before we
uh wrap up?
Jason Belk (37:40):
Sure, yeah.
So I I'm primarily on LinkedInand and X Twitter uh as Reno
Belk.
I have a blue sky, but I I Ionly have so many places I can
keep track of.
And so I'd say I'd say but andthen I'm you know, I'm also on a
number of different Discordsthere out there, but yeah, uh
like I'm on the R of NetworkExchange.
People want to ping me therefor that one.
(38:00):
Um, but uh of course I'd saylike the primary primary one
would be LinkedIn.
That's the one I check the mostoften for Reno Belk.
Reno Belk.
At Reno Belk.
Tim (38:08):
Yeah.
Chris (38:08):
We'll we'll make sure we have the links on the sky's
dying.
Um hate to say it, but youknow.
Tim (38:13):
Yeah, is it well?
I mean a lot of people that'sthe problem with the that's the
problem with moving uh peopleover, is that a certain center
of gravity has to move in orderfor the thing to work.
So anyway.
Yeah.
All right.
Um okay, well, yeah, let's goahead and uh wrap it there.
And we'll get all the links,and there's gonna be a lot of
links in this one.
Uh so definitely go through allof them.
(38:35):
And uh yeah, we will thanks forjoining us, Jason, and we'll um
see everybody on the next uhshow.
Sounds good.
Chris (38:42):
Thanks for guys.
See you guys.
Hello and welcome back to another episode of the Cables to
Clouds podcast.
And uh with me as always is myco-host, Chris Miles.
I am Tim McConaughey, and wehave a guest with us uh from
Learning at Cisco, uh JasonBelk.
Jason, go ahead and introduceyourself.
I don't think you've been onthe podcast before.
Jason Belk (00:31):
Hi.
Thanks, Tim.
Thanks, thanks, Chris.
Yeah, I am Jason Belk.
I am a senior technicaladvocate with Learned with
Cisco.
Um, the people who do thetraining and certification, CCNA
up to CCIE and everything inbetween.
And uh a little bit aboutmyself.
I live in Reno, Nevada, married12 years, have two young
daughters, a black lab, and uhmy background is in I've been at
(00:52):
Cisco since 2010, off and on.
And I I started out at Cisco ITas an analyst, network
engineer, network automation,started teaching networking
automation before DevNet to myour networking team, learning it
myself, and then I left andwent to network to code for a
few years.
Um, came back to Cisco as a indeveloper relations in the Cisco
NSO BU, which is like basicallya big automation product, and
(01:13):
then now more recently inlearning with Cisco.
So kind of done around thecircuit inside and outside of
Cisco and have some operationalchops to show for it.
So I kind of know the strugglesof all the operators out there,
uh at least for the on-premisestuff and the enterprise.
I was on the campus and branchteam when I was on call.
Um, so yeah, and and I'm reallythrilled to be here.
I mean, we've had conversationswith with you guys off and on
for a while now, and big fan ofthe show and love what you guys
(01:36):
do both for the news perspectiveas well as covering, you know,
all the different stuff fornetwork engineers who are trying
to grapple with AI and cloudand everything else.
Thanks, man.
Tim (01:45):
Um, so yeah, yeah, what's that?
Oh, I just said I appreciatethat.
Oh, appreciate that.
Yes, we are appreciative.
Uh it's always nice to meetsomeone who actually listens to
the show.
I'm just kidding.
Um anyway, so we broughtactually we brought Jason on uh
because he is working onsomething that's actually pretty
(02:05):
cool.
Uh we've been we've beentalking it back and forth for a
long time, actually.
We've been having trying to getuh Jason on the show, so we
finally made it work.
Um and Jason is here to talkabout something that I didn't
know existed, and I guess thatprobably a lot of the listeners
were not aware existed.
Uh something called spec drivendevelopment.
And uh I believe it hassomething to do with with
building for AI, but actually,Jason, I'll you're the expert.
(02:26):
I'll let you I'll let you takeit from here.
What is spec drivendevelopment?
Jason Belk (02:29):
Sure.
I mean, I've been using it forabout six weeks.
Uh and AI speak that makes mean expert.
Um that's right.
Uh to be fair, no, like one Iactually asked Claude Coder or
today how much I'd been using itjust on one project, and it
said 150 hours in the past threeweeks.
Um project, one one personalproject.
(02:51):
Um and I've and I I have likefive or six personal projects
going on top of some work stuff.
And and so all that to say iswhat is spectrum development?
I mean, if you think if you'vebeen in IT or any amount of
time, you're familiar withproject managers.
They're those friendly peoplewho help keep you on track and
get all your T's crossed andyour I's dotted.
And if you if you've tried tobuild anything of significance,
(03:13):
you know it helps to have aplan.
And so when you're talkingabout an AI, you you I think
we've all heard of the exampleof AI being, you know, like an
expert idiot or something.
Like they have tons ofknowledge, tons of ability to
execute, but what I would say isvery poor wisdom and judgment.
And so how do how do you helpsomebody who is able to do
whatever you say and sometimesinterprets it a little too in a
(03:37):
direction you weren't expecting?
Is that that you give them aplan.
And there's definitely thisisn't the only way of
approaching that problem.
And this isn't for um, youknow, I'm on chat GBT, Jim and
I, Claude, you know, I'm justdoing my chat bot and I'm
talking back and forth.
I'm not talking about thatscenario.
The the way that I'm I'mtalking about specifically is is
for a quote unquote agenticcoding, um, where where you are
(03:58):
giving prompts, instructions toyour computer through a text
interface, in my case, usually aterminal, and you're you're
typing in just as you would onChatGBT or Claude or Gemini, and
instead of getting a responseback that says, hey, I can't do
that, sorry, it actually goesoff and tries to do it using
your computer resources.
Um so you have to havepermissions and trust in the
(04:20):
place where it's running for itto execute that code.
And by default, something likeClaude Code, which is one I'll
be referencing in the rest ofthis conversation just for sake
of simplicity, um, will ask youpermissions when it wants to
create files, when it wants todo web searches, when it wants
to edit edit or delete files, dobash commands, stuff like that.
But all that to say is that itcan do all those things for you.
(04:40):
And not just from a codingperspective.
Um, there's another podcast Ireally enjoy called Software
Defined Talk.
And they had a whole episodeabout cloud code I'd recommend
people check out.
If you haven't heard about itbefore, played around with it
before, it gives you a goodprimer of what I would say
somebody who's not a coder or anactive coder using cloud code
just for productivity stuff,like sorting your screenshots in
your screenshots folder andnaming them, um, or uh, you
(05:02):
know, putting together yournotes and things like that.
Because it can basically onceyou give it access to those
files, and that's of course atrust issue, it will then be
able to do instructions based onrenaming, reading, and all that
kind of stuff.
So going back to spec drivendevelopment is that okay, you
you give this a local AI agentthe capability to do things on
your behalf.
And if you start asking to dothings, as I think we've people
(05:24):
who have worked with AI for awhile will experience there's a
context window issue.
And I'd say there's also like ascope issue.
And like I meant to do this,but actually you started doing
this other thing, and thenthat's not too bad, but actually
you were kind of deviating fromwhat I originally talked about,
or you built a production gradesystem, and I was just trying
to figure out if it worked.
Um, and or most common one Istill even run into, even with
(05:46):
spec driven developmentsometimes, is that you're trying
to pip install this on mysystem Python, and I want to use
UV.
Um, so spec driven development,not to say it solves every
possible problem, but I've gonefrom like having to intervene
and course correct my quad codemultiple times per session to
like maybe once a week.
And I can just let it run foranywhere between 30 seconds and
(06:08):
a minute.
I've had it sometimes runningup to several hours.
Chris (06:12):
So it's funny you mentioned that episode of
Software Defined Talk.
I just listened to that lastweek and it was uh it's it's
it's funny because like theythat was like my motivator was
like, man, I I really need todig in with clawed code because
they even call it out in theepisode, but like the fact that
it's called clawed code kind oflimits what you think in your
brain that you can do with itbecause you can do very vast
(06:34):
things with it.
But um, yeah, I like I likewhat you said about the the idea
of you know spec drivendevelopment kind of being that
specification that you define upfront to kind of keep your AI
in line.
But I'm I'm curious, like uh Ifeel like I've I've heard of
spec-driven development uh uhyou know several times,
obviously very recently as well.
Um, was this the thing thatjust came along with AI almost
(06:58):
as a not guardrails, but kind oflike guiding light for AI um
coding?
Or was this there before?
Jason Belk (07:05):
I'd say that the principles, um, and and I'm I'm
we'll share some links in theshow notes that people can take
a look at.
Um the principles I think goback to project management
principles in terms of definingrequirements and and and
executing executing upon them,uh, you know, clarifying
assumptions uh in a more formalway.
Uh, but I'd say, at least frommy knowledge of it, I've been
following it since August of2025.
(07:27):
I think that's around the timeit came out.
Um, because I I follow a lot ofdifferent YouTube channels.
One of them is like the GitHub,you know, AI folk, and and they
had some Microsoft DeveloperDay back then where they were
walking through it.
Um and so the the the one thatI use is the GitHub, what they
call spec kit, and it's onGitHub for free.
(07:47):
I'm I'm also sharing, um, I'llshare a link to my personal
spec-driven development kind offork of it, um, with the caveat
that the cut the code I have inthere that is the spec driven
development part is uh is notlike updating with the remote
fork.
I haven't got that complicatedon it yet.
So if you want to have thelatest GitHub spec kit part, you
might have to swap that partout and then use the rest of my
(08:08):
bits I have in there.
All that to say is um it's it'sstill relatively recent, as of
August.
And then I I tried to get itset up myself a couple times,
but I'd say the documentationout there for me as an
intermediate level programmer,novice level AI person who is
figuring stuff out as long as Ifor the past since 2023 or
whatever, since GPT 3.5.
(08:29):
Um I I I had a couple of stopsand starts until this guy named
Jason Gekki, who who's um I I'mconnected with who's people at
work pointed me to some umdocumentation that he does.
He basically he created apresentation that is happy to
share with whoever it wouldhelp.
And he's using this a lot, hisstartup and and consulting.
And so I'm gonna share the linkwith that as well, which I
think just helps visualize theflow and how all the pieces fit
(08:52):
together.
Because if you just go to theGitHub repo, and this is where I
got stuck, I'm like, okay, sowhat do I do with this?
I I see we have a repo and Isee I have clawed code, but
like, where do the dots connect?
Chris (09:01):
Yeah, for sure.
Jason Belk (09:02):
And and the first couple of times I tried, I kind
of gave up before I got throughhis stuff.
Um, not through it in a badway, but like before I I I went
through his his the way heexplained it, and then I've
tried to explain it as well onCisco U and a tutorial we'll
talk we can talk about a littlebit later, is just kind of like
connecting that dots of like,okay, how do I actually use
this?
And like it's not it's reallynot as mystical as it sounds.
It's and once you I I even wentthrough a demo of this with one
(09:26):
of my friends from church, andhe's he's like he has some ideas
on on a potential businessidea, and so I I shared my
screen and my laptop on thecomputer, and I literally did
the spec spec flow with him forhis idea.
He was defining therequirements verbally to me.
I would just type in whateverhe said, and then we literally
built a product in like 45minutes.
Tim (09:41):
So help me out here because um I'm new to the spec driven
thing.
Sure.
Uh the spectrum, so you know,I've been messing around with
cloud code, I've done somethings, I've failed at more
things probably than I'vesucceeded at.
Uh I tried to make a news agentfor the show and I posted that
about that on Substack.
Um if you know, me beingknowing as little as I do, but
(10:03):
and and knowing, you know, beingdangerous with it, I
supposedly.
How do I how would I, as aperson who's messing with it,
getting better but not great, umintegrate?
Or where do I even start tryingto pull in spec driven
development into like my ownworkflow to build stuff?
Jason Belk (10:19):
Sure.
I I would say start with agreenfield project, something
that like you had an idea.
It's you're not trying to boilthe ocean, like I need I need to
make the next Facebook app orsomething.
But if you're an infrastructureengineer, you're like, okay,
maybe I I'm not supercomfortable with Ansible and I
want to deploy some Terraform aswell.
And like, okay, uh, or I knowTerraform really well and I
don't know Ansible.
And so it's like you have anyou have a context area that you
(10:40):
already kind of understand, butit's kind of rusty and you have
to look up a lot of docs.
So you're able to define therequirements reasonably well.
So I'd say, like, likebrainstorm, okay, this is what I
kind of want to do.
And you don't have to get itall right in the first shot.
This is more just kind of liketo make sure set you up to
succeed.
Now I'd say check out the linksthat we were we set up because
it's it's something that I'lltalk through, but it's almost
more important for you to lookthrough and do it yourself with
(11:01):
your fingers on the keyboard.
Um, because it it's it's reallynot rocket science.
Like once you have on onescreen the flow and the other
screen your terminal, and likeyou have it set up.
Okay.
The the installation process isbasically just a bunch of bash
scripts.
It's not like installing anynew software, it can kind of run
independently too, just as it'sa bunch of markdown files and
bash scripts that kind of likeguide the quad code to do the
(11:23):
the project management workflow.
And so the the the GitHub repothat I have for for my um kind
of additional context, I'll justsay some additional
instructions on the spec kitflow, which is optional, or the
core one that you can get fromthe official Microsoft GitHub.
Um you can basically get clonedown the spec kit.
I'd say the one thing I learnedthe first time I did that I did
it incorrectly is that Iinitialized my clawed code at
(11:44):
the wrong directory because Iwas kind of new to Claude Code,
where I was like, okay, uh I'm Ihave my project folder, make
dira project, and then I putspec kit in that in that
subfolder, and then I ran clawedcode at that layer above where
the spec kit was installed.
So it didn't look because theseare technically what are called
claud skills, and it doesn't itdoesn't load in the skill if
you're not in that directorywhen you initialize claude code,
(12:05):
which it didn't tell me.
And I mean like I I went allthat to say is like if you're
doing it, just make sureinitialized when you type in
claude on your terminal when youhave it installed, do it in the
same place, the same layer ofthe directory where you have the
spec kit installed so it canload in those skills.
Tim (12:19):
Oh, okay, I see.
So so I mean you're pullingdown a repo, like you said,
basically, and and and in thatrepo is gonna be like skill
file, like files essentiallythat that define all the skills
for that you're gonna use forthe for the workflow that you're
gonna follow on the spec drivendevelopment.
Okay, that that makes it alittle bit more real to
visualize.
Okay, I see.
Jason Belk (12:38):
And you can inspect them, and it's not like some
binary that you can't evertouch.
You can look at the bashscripts, you can look at the the
markdown files that are a partof it.
Um but basically as soon as youload up that clawed instance
and it loads in those skills,you can then access on slash
commands so you to to load it touse those skills.
So you go, you know, like slashspec kit dot, and then the
first one would be constitution.
(12:59):
And if you do it in the wrongorder, it'll kind of like gently
correct you.
It's like, oh, you're trying toimplement, you don't have a
plan yet.
Tim (13:05):
You want to have a plan it's a workflow, it's yeah, it's
a whole it's like uh uhwaterfall.
Yeah, yeah.
Okay, yeah, that makes sense.
Jason Belk (13:13):
Yep.
So so the the the first step,and this is one that I'd say it
doesn't always enforce, but it'sgood to start with because it
kind of gets you in the rightheadspace and helps it up
farther down the road, is theconstitution.
So just like in the UnitedStates and other countries, they
have a governing body thatdefines the principles of what
should be executed.
Uh so in in this case, it'sbetween I think three and five.
(13:34):
Generally, it suggests.
And and it's not that you haveto walk in with like a quill and
like some parchment and and andand have it all worked out
before you start.
Like Hancock on there.
Chris (13:44):
Yeah, get your drawing Hancock.
Jason Belk (13:46):
That's right.
I actually use um this is kindof like a JSON hack that I
started recently, one of myfriends pointed me to as Whisper
Flow.
This is different than spec hitdevelopment, WISPR flow.
It allows you to do voice totext real easily.
So you have to have a hotkey,and you can do that prompting
into your terminal.
So it makes my prompts so muchbetter on Claud Code, where I'm
not like getting carpal tunnelDM VPN on my on my Claud Code.
(14:11):
I just hold down option, that'sthe key I set up, and I just
talk for like 30 or 45 seconds,brainstorming what I want to do
at the consultation at theconstitution where it'd be more
at first principles like, okay,yeah, I'm trying to do a in this
case Ansible Terraform app.
I kind of I know Ansible reallywell, but I don't know
Terraform so well.
Um my my resources are allgonna be public cloud, AWS,
Azure, and like my keys aregonna be in this.m file or in my
(14:34):
you know my my ZH S HRC and allthat kind of stuff.
So you just kind of like giveit the lay of the land and kind
of what direct what whatdirection you're heading.
And I just kind of record 34 to5 seconds, paste it in there,
and I'm done.
And and then it processesthrough there and uses the the
LLM.
And that's a one other keything for me is not to use
sonnet, but to use Opus.
(14:54):
When I did spec drivendevelopment, like it default.
I mean, I think quad codedefaults to Sonnet, which is
fine for like you know settingsimple stuff, but you can do
slash model and switch to opus.
And yeah, you're gonna use upyour tokens a lot faster, but
it's like cry once, try once, orwhatever, and like because
otherwise it was just spinningin circles with Sonnet.
It's like, oh sorry, I brokesomething.
Let me fix it.
(15:15):
Oh, I broke something else.
Let me fix it.
And so I was like, with Opus,it still breaks things on
occasion, like once a day forme, or and what stuff like that.
But I'd just say, even if youhave less Opus tokens based on
your pro or whatever level, justwait till it refreshes.
Yeah.
Tim (15:30):
Yeah.
So this is interesting becausewhen I was working on my agent,
um, I was trying to save savetokens.
So like I was using um I usedhaiku, which I was like, oh wow,
this is a really bad idea.
It's a really bad idea.
I realized it after the fact,after I was building the code
for it, I was like, wait aminute, I shouldn't use haiku to
build this.
But what haiku is good at iswriting.
(15:50):
So maybe from like this part ofit, this constitution defining
like doing the writing part ofit, would that be better just to
get it all in the mod?
Like get it because all you'redoing at this point with spec I
think, and correct me, is atthis point, really what you're
doing is creating the files thatclause refer to uh refer to for
(16:11):
the development part of it,right?
Jason Belk (16:12):
Yeah, yeah, I'd say I haven't tried that.
It's a good idea.
Like in terms of the flow, I'dsay definitely when you get to
the implementation layer, youwant to switch to over.
But you could probably get awaywith it on Sonic and Haiku on
the on the initial ones.
I I haven't done like A-Btesting to see what the
difference is, but I think formost, I think on the
implementation level, that'sdefinitely for your own.
Tim (16:29):
Oh, it's definitely no doubt, no doubt.
I can tell you haiku isterrible at implementation, not
accident.
I can tell you that for sure.
Jason Belk (16:35):
That's a good idea.
Um, and so yeah, for once youget your constitution.
So I I do my brain dump, itresponds back basically, usually
like, hey, I've noticed basedon what you said, or if you
didn't give enough information,I'll ask them follow up
questions.
Um, based on what you said,here's three to five principles
that I'm seeing.
Like that you you want to havesimplicity as a design, you want
(16:56):
to have cloud native as adesign, you want to have these
you're doing this as a proof ofconcept rather than a production
build.
So like it kind of extracts outor clarifies for you what type
of project you're trying tobuild.
And this doesn't just have tobe for coding.
Um, this could be for for otherstuff too.
Like that's but all all that tosay is like what then it
creates from the that response,then it generates your
(17:16):
constitution file for you.
So I'm not like going in therehandcrafting.
You can you can go in there andedit the file, but it basically
writes it for you.
And then you go to specify.
So once you have yourconstitution going, you flow
through through these this loopof this workflow for each of
your features.
So you go from specify to planto tasks to implement, back to
specify to plan to tasks toimplement for each of your
(17:39):
features, basically.
Um, and and and I've found forbigger projects, oftentimes my
first spec is like anarchitectural one where its
output is not code, but justlike architectural documents,
um, where I'm like, okay, let'sspecify what cloud we're gonna
be using, what you know, whatlogging we're gonna be using,
what um all this kind of stuff.
Where uh if it's a much bigger,if you know it's gonna be a big
(17:59):
project, that'd be myrecommendation because sometimes
it just it just makes decisionsand does even with the spec kit
flow is gonna start making someassumptions on things that you
might have to go back and reworklater, which isn't in the
world, but like that'd be my onerecommendation for a bigger
project.
But if you're just gettingstarted doing a small thing like
this, it doesn't matter.
I would say you're probablyjust gonna have maybe two or
three specs tops.
Um so the first the first onewould be, you know, like the
first one that flows.
So we have the constitution,and then you go to and then you
(18:21):
do the slash command slash speckit, and it'll have like tab
complete if that skill is youknow um loaded in there
correctly.
And then it sometimes reallyit'll even suggest, like at the
end of its constitution, like,okay, we have the constitution
already.
Do you are you ready for speckit specify and like copy and
paste?
Yep.
Um and and and so then for speckit specify, you're documenting
your requirements.
So going back to that projectmanagement example, you're
(18:43):
sitting down with the projectmanager, and they're like, Okay,
so what's what's the problem?
Like, well, we have a new salesoffice and we need to get
access in for salesforce.com.
And they we we have an appwe'll have to take down the
network over the weekend inorder for this this network
change to take place.
Um, and no, we're gonna do iton the Cisco iOS cat9 gates and
(19:04):
Cisco Catalyst Center orsomething.
So you're just defining youryour requirements in terms of
what you're trying to accomplishfor this particular feature.
Tim (19:12):
Yeah, that makes sense.
I mean, you're setting up theguard essentially the
guardrails, like what are yougoing to be working with?
What are you gonna be doing,right?
Chris (19:19):
Like I think uh I think one thing that I've definitely
observed is it's amazing to mehow fast Claude and Anthropic
have kind of switched to thefront runner in this in this
conversation.
Um it's it's been you knowspecifically over the last it
seems like over the last monthlike it's really really picked
(19:39):
up steam.
Um but one thing I've alsonoticed is the the kind of you
know kind of variety of toolsthat people are using.
Like you said, you're usingsomething specifically for voice
to text, and there'sintegrations and there's
people's building skills andusing different MCPs and things
like that.
So maybe can we can we take astep back and let's talk a
little bit about skills andbecause I know with with spec
(20:02):
driven development, um it seemsone of the kind of concrete
things is is the idea of becauseyou've well defined your
requirements and your intent upfront, this allows for kind of
parallel development, right?
You have this you have agentsworking on one piece of it and
agents working on another pieceof it, and they can operate
autonomously, right?
(20:23):
And they won't conflict witheach other because they all are
working off of the samerequirement, right?
Correct.
Um, so can you tell me a littlebit how agents and skills play
into that?
Jason Belk (20:32):
Yeah, I I I'm by no means an expert on um skills,
but I'll my humble explanationis that they're just a series of
bash scripts and markdownfiles, primarily markdown files,
that um help the model, usuallyClaude, understand what parts
of its training and brain to usein that particular context.
So just like if you're onChatGBT and you say, pretend
(20:53):
you're an expert networkengineer and analyze this ACL.
Okay.
Um it's it's it's it's it'sit's it's funny because like on
the software-defined talkpodcast, they actually even
talked about like if you if youlook at like the market share
that's been destroyed by a bunchof markdown files, they're
like, it's tremendous.
They're like, you know, you'redestroying accounting, it's
destroying legal.
It's like you look at whatactually is destroying, it's
just it's just like somemarkdown files on these repos.
(21:15):
It's not like they they didanything super special, it's
just like it it's such apowerful model, it just needs
like little guardrails and likesome to say, okay, just focus on
this part and take on thisidentity.
Um yeah.
Tim (21:29):
So actually, real quick, because uh you just reminded me
of something, and I I this ismore of an opinion thing because
like I don't think anybodyknows what's gonna happen in the
next six months or three monthsor two months with AI, these
things.
But isn't this this feels likeuh, you know, what I'm hearing
more and what I truly believeprobably will be the the result,
the end result here is as weseek to make these things more
(21:49):
efficient, and as trainingbecomes things that, you know,
we have enough material uhhardware to start doing more
training, specialized training.
Do you think that we're gonnagravitate towards more like
small, you know.
Small specialized models thatare going to be built kind of
for this, or unless, or do youthink it's going to be like, oh,
well, here's an agent, we'regoing to put the guardrails in
place, and then we're just goingto have the LOM still back at
(22:10):
all?
Jason Belk (22:11):
I think it's going to, my opinion, just the way
I've seen other stuff work,you're not going to know or care
as a user.
Sure, no, 100%.
But but but like from atechnical perspective, um Yeah,
that's what I'm thinking.
I I I I can definitely see thatbeing in the approach.
I mean, that's why I work atCisco, like they released a
security um L um special trainedmodel.
Yeah, a small language modelfor that last June on Hugging
(22:34):
Face.
Because uh yeah, like I thinkit's also just not feasible to
maintain these models.
I mean, I might be wrong, I'mnot an expert by any means, to
like try to pretend like thecorpus of every possible job in
existence that will exist andthat will exist, could be
maintained by something that orwould want to be maintained,
even if we could, it's like,well, we could we could use all
(22:55):
of the power from Idaho to likepower this thing, or or or we
could just run it on your run onyour phone.
Tim (22:59):
And so that's kind of my feeling too.
Like, so I I think that'sprobably where we're going to be
headed once once we get more ofthe of the process of being
able to truly kick crank outthese small language models or
specialized language models,depending on how you want to
talk about it.
Jason Belk (23:14):
Yeah, I think it's right now such a new thing for
so many people and so disruptiveto people's workflows and such
an open form thing.
Or previously, with like cloud,you're like, okay, I have to
get like my AWS keys and likeyou know, go onto the console
and like figure out what I wantto do here, or or even from a
SaaS perspective, like okay, Ineed to like get my credit card
and like get this SaaS thingplugged into my work workflow.
(23:35):
And that there was still likesome complexity there and like
some changes to people'sexisting paradigm.
They're like on-prem versuscloud, hybrid, you know, all
that kind of stuff versus likeLMs just blows everything out of
the water.
I was like, what if you couldjust like talk to your computer
like in Star Trek and it couldlike answer?
Right.
And it's just like I I I Imean, I'm not trying to like
trivialize anything about this,but I think that's in my mind
(23:57):
one of the foundationaldifferences that it's so open
form and free form that it'shard to train people on like
what to do.
You're like, oh, I you mean Ican have it help me cook dinner
or build out ACLs on my router.
Okay.
Um, but I guess we'll do that.
Um, so I I I think that's mymind what the adoption curve is
(24:18):
gonna is is is gonna eventuallylook like is that it's gonna get
more adopted as you get morespecialized bottles where people
can understand their use case.
They're like, oh, this is mysecurity model, I'll ask it
security questions, even thoughthey could probably do that
right now.
Tim (24:29):
Um yeah, yeah, yeah.
But efficiency-wise, of course,you gotta when you every time
the ask and LM a question,you're lighting up like half the
planet to to get the answer.
Um you said something earlieralso about the um oh about the
uh oh man, I just lost my trainof thought, of course.
Darn it, I I had it for asecond and then I just lost it.
(24:50):
I give up.
Uh Chris, uh I'll head it overto you.
I was I was kidding.
Jason Belk (24:54):
I was I was put I was put in the spot for no
reason for that.
Tim (24:57):
Sorry.
Jason Belk (24:58):
I I can keep going through this stuff that's uh so
so we we we coveredconstitution, we covered
specify.
So after specify, you have theplanning phase.
So you've defined yourrequirements, you've defined
your governing principles, andthen now you have your
requirements, you go per featuredesigning your technical
approach.
And and and this would be justthe next flash command.
So after it does the specify,and that might take several
(25:20):
minutes or longer in terms ofthe specifying the constitution.
So you gotta be patient, go tothe bathroom, take the dog for a
walk, um, and then come backand then do the spec kit plan.
And you can do these inmultiple tabs.
Like I'll sometimes have threeor five different spec kits
running on different projects atthe same time, which just makes
you feel like superhuman.
Um and so, like spec kit planper feature design technical
(25:42):
approach, and that that'sbasically like, okay, now that
we have our requirements, likewhat's gonna be our approach for
the Ansible Terraform example?
It's like, okay, you know, likewhat modules are we gonna use
or for Ansible and and what umproviders are we gonna use for
Terraform?
And this isn't something thatlike you have to define that out
because like it based on theLM's knowledge, and if it has
access to the internet, it canfigure that stuff out.
(26:02):
So, like this plan isgenerally, in my experience,
generated from the your maininput is on the specify, and
then the plan it builds a planfor you.
You're not like okay, I have tospend half an hour figuring out
my plan.
It's like, hey, based on yourrequirements, this is why I
think a good plan would be basedon your governing principles
from your constitution.
And you're like, that is a goodplan.
Let's do that one, or you cango in there and you can clarify,
(26:23):
and this is for the voice totext input, I'd be like, I'd be
like, no, actually, like Idecided not to use Ansible.
We're we're we're gonna usePuppet and Chef instead.
And it'd be like, okay, I guesswe need to go, we we need to
rechange everything.
And so, like, you you can stillpart way through change your
mind and do different things,but or maybe a more realistic
change to the plan would belike, okay, we're gonna have
this on specific jump posts thatyou need to know about and and
(26:44):
and and stuff like that, havingmore details on your on on what
is gonna affect the plan.
Tim (26:49):
Actually, you just reminded me, just reminding me what I
lost my frame of thought on,which is what we were talking
about, and and I think everybodywho's probably now put any
amount of time into AI isfiguring has already figured
this out, but just for settingpurposes, um, the more you know
about the technology you want touse, the easier it is to get
the AI to to under because ofcourse, because you can say,
(27:12):
like, oh, I know you're gonnaneed this information, right?
I know you're gonna need toknow about this, I know you're
gonna end up like if you knowAnsible, like if you know enough
about Ansible, you know exactlywhat that you gotta tell make
makes it into the spec.
Yeah, you you know exactly whatneeds to go in there.
Jason Belk (27:26):
Yeah, yeah, you know what the gotchas are.
And so this is where I thinkpeople who don't have a
technical background use AI, dokind of get caught up or get
stuck or get in rabbit holes.
And so this is like exactly toyour point, like your advantage
as a technologist in definingthese things is that you can
front load your assumptions onall the things you wish, as if
you were talking to yourselffrom a mentor version of
(27:48):
yourself where you've forgotteneverything, but you also you're
like this is yeah, or a junior,a junior engineer, and so you
need to clarify yourassumptions.
And and this is kind of mything that I've realized over
the past year, year and a halfis that like your leadership
skills, your soft skills as ahuman being, make your ability
to execute prompting andplanning like this uh
significantly better.
(28:09):
So I'd say brush up on somebooks on leadership if you have
nothing to do and you're boredon the technology side, and it's
gonna help you be a better AIuser in the future because it
gives you things like clarifyingassumptions, you know, defining
requirements and all thesethings that like project
managers and people with withyou know technical leadership
(28:30):
backgrounds are all used to orconsultants oftentimes too,
where you have to sit down anddefine your workflows.
Um it's uh it's all in yourhead somewhere where you you
need to get it on paper, kind ofassuming nothing.
Chris (28:40):
Yeah, I think um one thing I've definitely noticed in
working with AI forspecifically this type of uh
this type of use case is uh I'mvery good at giving context, and
I often give a very largeamount of context, which as you
guys talked about, it really uhit really messes with your you
(29:00):
know your windowing uhavailability when you're
interacting with AI, so it'slike burnt me a few times
because I'm like I'm like no,this is important, you need to
know this.
So I do like this idea.
Yeah, I do like this idea ofbeing able to define that up
front and that just using thatas a reference source over and
over again.
Um because yeah, I mean, butthen there's tools as well out
(29:21):
there, like it we should weshould put a link to that
software defined talk episode inthe show notes because it's so
valuable to me.
Um but there's um theyreferenced that there's even
tools out there that you can usethat actually remove all the
fluff from any of your promptsthat like can't will not, you
know, so that you're not eatingup tokens that's cost.
(29:42):
You might be saving them.
Yeah, exactly.
Like you might be using themuh, you know, maybe 10% less,
but that 10% can be you knowthat's valuable for you down the
down the road, right?
Jason Belk (29:52):
And I'd say if you have other documents, you can
reference those as long asyou're you're willing to give
access to those files to thissystem.
Right.
So like for one of my projects,I had done some work in a
planning phase, not even beforeI even knew about Spectre
Development.
I I had like a 20-pageMicrosoft Word document that I I
had defined basicallyrequirements.
It was the first half of it wasgoverning principles, and the
(30:12):
second half of it was executionlines, not like code, but more
like defining tasks and things.
Yeah.
Um, and I p I found PDFs to bethe longer franca of of AI, and
so I exported the doc.
I think you can read doc X, butI always just do it as a PDF
just to make sure theformatting's right.
So I export as a PDF, includedthat in my constitution
specifying that like it wasreally good at that, because
(30:34):
it's like it it extracted allthe important details and kind
of like filled out the stuff forme on that stuff.
So if you already have aproject plan from like an actual
project manager, and they'relike, okay, okay, Chris, we need
we need you to to build outthis cloud deployment.
Here here's your here's yourproject plan.
You're like, okay, I'll exportas a PDF, and if my company lets
me use these types of tools, Icould just use spectrum
development to extract out andbuild out the plan from their
(30:57):
plan.
Tim (30:58):
Yeah, that makes sense.
And I don't run into that.
Jason Belk (31:01):
Oh, go ahead.
I was just saying once so onceyou one thing is that it
sometimes your plan might haveambiguities where e either um
based on what you're trying toimplement or the different
technologies and outcomes you'retrying to achieve.
And I'm trying to think of agood one.
Um, but an example might be ifyou weren't specific on a cloud.
If you're like, okay, I want todo Ansible and Terraform, but
you never really said if it wasAWS, Azure, or something else.
(31:24):
And it will oftentimesauto-suggest in at the end of
its response, like, do you wantto go to the next phase or do
you want to do spec kit forclarify?
And you can still run it, evenif it doesn't suggest it, and it
might say, Nope, there'snothing just to clarify, but uh
sometimes it will it willactually go through a clarify
phase where it goes through allthe tasks, I mean all the plan
that it's built and say, Here'smy plan.
(31:44):
There are actually someambiguities on some of these
plans, and then it will it'll dobetween one and five questions
that are multiple choice, or youcan fill in your own answer and
and it and and allows you toclarify.
So it's like like, for example,those questions you'll say,
like, I noticed you know ourplan we're we're we're we're not
saying which cloud we're using.
Do you and it'll kind of knowbecause it's smart in terms of
(32:06):
like, did you mean to use AWS A,B, Azure, C, you know, GCP, or
D, something else?
Fill it in.
And it doesn't have to be aone-word answer.
You could I can say, well, I II I have an Azure account, but I
prefer AWS, and so it'll belike, okay, no, let me update
the plan.
Tim (32:20):
Yeah.
I like to use the uh I mean, Ithis is uh obviously spec driven
is much more uh not justcomplicated, not just complex,
but obviously very, I assume inthe weeds with it.
But I mean I've been using theplanning mode of Claude quite a
bit.
Um for the it for the same forthe same reason, right?
Create me a file because I knowwhat's gonna happen is that I'm
(32:42):
gonna run out of context whenyou can compact the
conversation.
I'm gonna lose all the detailsI put in.
Exactly.
Go back and refer to this fileagain when when you've when you
after you've done that, right?
Jason Belk (32:52):
Exactly.
And that's one thing Imentioned early on, but I forgot
to mention again, is that thatis the huge advantage of this or
the planning mode on ClaudeCode is that once you get to a
certain point, all its tokens inits brain get scrunched down to
a little dot, and and then ityou you kind of have to start
over again to certain steps.
But when you have this plan allwritten out, it just refers
back to the files.
It kind of knows where it is inthe flow, it has like its own
checklist.
Like, okay, I'm on the planningphase.
(33:12):
Now let's move on to the tasksphase.
Tim (33:14):
Yeah, it's a it's a big big bonus.
Um I I yeah, I had to learnabout it uh because I didn't
realize that until I did threeor four projects and I was
losing all my context and it wasforgetting everything.
And I was like, man, there'sgonna be a better way.
And so yeah, that's the anyway.
No, this is good.
Um we've got all the so we haveall of the the links, including
(33:35):
the one for uh that goesthrough the entire process of
spectrum development.
I think that's gonna be use uhreally useful.
I think we just that I thinkthat's probably the first thing
where people what I'm thinkingabout is okay, there's two
things, right?
One is where do people start?
We kind of covered that, whichis like literally just go look
at these documents and then gofork a repo or grab download the
repo.
That's so that's just to thelisteners that are already work
(33:57):
with AI or that want to startworking with AI and want more of
a framework to start from.
That's the first one.
Um, the second one is uh sincewe have to we'll we'll start uh
wrapping up here in a minute,but let's talk a little bit just
before we wrap about the CiscoU cores that you built.
What's at like compared to theto the docs that are the the
(34:17):
docs we were like to see?
What is the what covers what'scovered in the course?
Jason Belk (34:20):
So so yeah, I we have a couple links here.
We have one that's a tutorial,these are free.
You don't need anysubscription, you just need a
CCO Cisco account, which doesn'teven have to be a Cisco
customer.
You just you can create one forfree with your Gmail, I think,
or whatever.
And and so this to Cisco Ututorial, Cisco U is our
learning and certification,learning with Cisco, place where
you get your certificationtraining officially from Cisco.
(34:40):
So they have you know CCNA upto I think CCNP and and other
training in between.
Um and so we also have freetutorials on a variety of topics
from cloud, security, AI,networking, the gamut.
Uh I think we have over 200 atthis point.
So not just this one tutorial,but check out the other ones.
We have ones linked todifferent certifications you
might be pursuing as well.
So for for this particular oneI wrote a few weeks ago, I
(35:03):
basically took the spec drivendevelopment um gecky link that I
share, I talked about earlierand kind of converted that to a
network engineering context.
So if you uh if you look at hisgreat presentation link and
you're like, okay, I kind of getwhat it is, but how do I do
that as a network engineer?
Look at my link on the Cisco U1where I basically have you go
through a spec driven flow forsetting up the skill and then
(35:23):
using it to like basically buildlike a simple Ginja 2 YAML
parser thing that creates someconfig files.
Um, so that should get you upand running, at least with a
very simple use case that's allgonna be run locally.
You don't need any remotedevices, you can just kind of do
it with your Claud code locallyon your personal laptop or
whatever to get started.
Uh and then I'd say on top ofthat, so that's the free one on
the paid side, and I think theymight have it free right now,
depending on when this getspublished, of a Cisco AI
(35:45):
technical practitioner course.
And we came out with acertification last December.
So, this in terms of complexitylevel, on those who are
familiar with Ciscocertifications, the CCNA Cisco
certified network associate isan associate level exam.
This is one step before that.
So this is meant to be entrylevel.
I don't know what I'm doingquite yet.
Let me figure it out, type ofcourse and exam.
So we came out with a courseand exam in December.
(36:06):
So it's been out for a fewmonths now, and it it covers
prompt engineering, you know, AIsecurity, AI-driven software
engineering, um, advancedprompting strategies, uh, you
know, what is an LLM, what isRAG, and all that kind of stuff.
And I'm I'm actually doing aCisco Live session covering this
AI tech course andcertification.
So if you're coming to CiscoLive US, US, come check out my
(36:27):
breakout session with withGeorge.
Nice an hour, it'd be a lot offun.
Um and then I'm also doinganother um Cisco Live session on
the new um CCMP automation,autocore, EN core, like the
enterprise networkingautomation, professional level
that used to be called DevNetProfessional Exams.
So I have another one therethat people should check out.
We're running Cisco Live, andthen third, I have what's called
(36:48):
Cisco U session.
So we have 45-minute sessionson on the floor, and that one
will be on Spec Drivendevelopment specifically.
So if you're still unsure andyou happen to be a Cisco Live,
check out my session there orthe recording will be up, you
know, after Cisco Live as well.
Awesome.
Chris (37:01):
Yeah, we'll definitely get all those shows or all those
links in the show notes aswell.
I think this will definitelypublish before Cisco Live.
So yeah, for sure.
If you are if you are going toCisco Live US, make sure you do
check out the session.
That sounds great.
Is it our details?
Do you have details of thesession as well?
We can put the I can get this,yeah.
Yeah, let me pop.
I can pop the side.
Tim (37:20):
Yeah, the catalog is up, catalogs up now, so we should be
able to we'll we'll get that.
Um I'm doing uh I'm doing alab, but I'm not here to plug.
I'll plug my stuff on adifferent show.
But uh yeah, so we'll get we'llget that.
We'll get that in the shownotes.
Uh definitely sign up for allof Jason's stuff.
Um also, Jason, where canpeople find you online before we
uh wrap up?
Jason Belk (37:40):
Sure, yeah.
So I I'm primarily on LinkedInand and X Twitter uh as Reno
Belk.
I have a blue sky, but I I Ionly have so many places I can
keep track of.
And so I'd say I'd say but andthen I'm you know, I'm also on a
number of different Discordsthere out there, but yeah, uh
like I'm on the R of NetworkExchange.
People want to ping me therefor that one.
(38:00):
Um, but uh of course I'd saylike the primary primary one
would be LinkedIn.
That's the one I check the mostoften for Reno Belk.
Reno Belk.
At Reno Belk.
Tim (38:08):
Yeah.
Chris (38:08):
We'll we'll make sure we have the links on the sky's
dying.
Um hate to say it, but youknow.
Tim (38:13):
Yeah, is it well?
I mean a lot of people that'sthe problem with the that's the
problem with moving uh peopleover, is that a certain center
of gravity has to move in orderfor the thing to work.
So anyway.
Yeah.
All right.
Um okay, well, yeah, let's goahead and uh wrap it there.
And we'll get all the links,and there's gonna be a lot of
links in this one.
Uh so definitely go through allof them.
(38:35):
And uh yeah, we will thanks forjoining us, Jason, and we'll um
see everybody on the next uhshow.
Sounds good.
Chris (38:42):
Thanks for guys.
See you guys.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by Audiochuck Media Company.
NFL Daily with Gregg Rosenthal
Gregg Rosenthal and a rotating crew of NFL Media hosts including Jourdan Rodrigue, Colleen Wolfe, and Nick Shook provide all the NFL news, previews, recaps and analysis you need to be smarter and funnier than your friends.