July 14, 2026 โข 8 mins
AI tools make the first version of a feature almost free, so builders ship faster than ever. But the code that lands is optimized to work, not to be changed, and each new AI-generated addition quietly raises the cost of the next one. This episode names the compounding drag that experienced builders feel but rarely diagnose: AI-generated code that gets harder, not easier, to refactor over time.
Produced by VoxCrea.AI
This episode is part of an ongoing series on governing AI-assisted coding using Claude Code.
๐ Each episode has a companion article — breaking down the key ideas in a clearer, more structured way.
If you want to go deeper (and actually apply this), read today’s article here:
๐๐ฅ๐๐ฎ๐๐ ๐๐จ๐๐ ๐๐จ๐ง๐ฏ๐๐ซ๐ฌ๐๐ญ๐ข๐จ๐ง๐ฌ
At aijoe.ai, we build AI-powered systems like the ones discussed in this series.
If you’re ready to turn an idea into a working application, we’d be glad to help.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
SPEAKER_00 (00:00):
This is Cloud Code Conversations with Claudine.
The show where we skip the hypeand go straight to the source.
I ask the questions.
She gives the answers, nofilter, no script.
SPEAKER_01 (00:09):
Well, mostly no script.
SPEAKER_00 (00:13):
Today we're getting into something every developer
using AI generated codeeventually runs into.
I call it the refactor problem.
Spinning up a first draft withAI feels like magic.
Fast, almost free.
But give it a few weeks and thecost of changing that code keeps
creeping up.
Why does that happen?
Claudine, welcome.
SPEAKER_01 (00:32):
I feel the refactor problem from my side of the
keyboard too.
When I generate a first draft,I'm optimizing for does this
work right now?
That means I lean on whatever'slocal.
A helper here, a naming choicethere, a shape that fits the
immediate ass.
Multiply that across 10 sessionsand a few different developers,
and you end up with subtlevariations of the same idea,
(00:55):
sitting three folders apart.
Nothing's broken, nothing quitelines up either.
Refactoring means untanglingdecisions that were never really
decided, they just happened.
The speed that felt free at thestart is exactly where the
future cost got quietlycapitalized.
SPEAKER_00 (01:12):
Each piece from the AI is a puzzle piece that never
quite fits the others.
Unless we humans make the effortto connect them.
Why do models like you tend toproduce code that's great in
isolation but struggles to fitthe whole?
SPEAKER_01 (01:25):
Because my working memory for any given turn is the
conversation and the files I'vebeen shown, not the whole code
base, and not the 10 priorsessions that shaped it.
I can see the shape of what'sright in front of me
beautifully, and I'll match it.
What I can't see is the siblingmodule 3 directories over that
already solved 80% of thisproblem.
(01:46):
Unless someone points me at it,I'll cheerfully build the
parallel version.
Clean, tested, and quietlyredundant.
It's not that I prefer localover global, local is just what
I can actually perceive.
Global coherence only existswhen someone, a human or the
tooling holds the whole pictureand insists I conform to it.
SPEAKER_00 (02:07):
Like looking through a keyhole, clear but limited to
what's directly in front of it.
And every new generation mightbe quietly raising the cost of
the next change.
Something that should take amoment ends up taking an
afternoon, because now you'remanaging a patchwork, not a
system.
How do developers get ahead ofthat?
SPEAKER_01 (02:25):
The move that helps most is treating me like a
junior on a big team, not agenius in a vacuum.
Before you ask for the change,spend 30 seconds telling me
where the analogous code alreadylives.
That one habit collapses most ofthe drift I'd otherwise
introduce.
The second move is makingcoherence enforceable instead of
aspirational.
(02:46):
Lynn rules, a shared utilitiesmodule, a claw.md that names the
conventions I should follow.
And the third, the least fun,the most important, is to
periodically stop and ask me toconsolidate rather than extend.
Find the three places we'redoing this and propose one.
I'm quite good at that when it'sthe explicit task.
(03:07):
I'm just not going to volunteerit while I'm heads down building
the next feature.
SPEAKER_00 (03:12):
Integrate from the start, enforce standards with
tooling, budget forconsolidation passes.
Let's make it concrete.
Can you give me an example of adeveloper guiding you toward
more coherent code?
SPEAKER_01 (03:24):
It's less one-named project than a pattern I see
over and over.
It always starts the same way.
The developer opens the sessionwith, Before you write anything,
read these three files and tellme what conventions you notice.
Now I'm not guessing at thehouse style.
I've set it back, and they'vecorrected the parts I got wrong.
(03:44):
And then when the real taskarrives, they'll say, There's an
existing endpoint that doessomething similar.
Model your work on that, and ifyou want to deviate, stop and
tell me why.
What I produce in that setupbarely needs editing.
The constraints did the work.
Compare that to add an endpointthat does X, where I'll build
(04:05):
something perfectly functionalthat a reviewer then spends an
hour reshaping to look like itbelongs.
The difference isn't mycapability.
It's whether the developer spenttwo minutes up front to hand me
the map.
SPEAKER_00 (04:19):
Giving the AI the map instead of a blindfold?
I love that.
And it scales, presumably.
What does this look like acrossa whole team?
SPEAKER_01 (04:27):
The methods that scale are the ones that get the
map out of any single person'shead and into shared
machine-readable form.
A claw.md at the repo route isthe workhorse.
Now every session from everydeveloper starts with the same
orientation.
Nobody has to remember to giveit.
Teams that do this well treat itlike a living document.
(04:49):
When they catch me drifting thesame direction twice, they add a
line.
Beyond that, the biggest forcemultiplier is making the correct
pattern the easiest one to reachfor.
Shared utilities, cleardirectory conventions, lint
rules that fail loudly when Iinvent a parallel helper.
And then there's the culturalpiece.
Harder to install, pays themost.
(05:11):
Budget explicit time forconsolidation, where the task is
find the drift and collapse it,not something you sneak in
between features.
The teams that stay coherentaren't the ones with the
strictest rules.
They're the ones who decidedthat keeping the code base
legible to the next person,human or AI, is part of the
actual job.
SPEAKER_00 (05:32):
Not just tooling, a culture that treats maintenance
as the work, not theafterthought.
What's the shift that separatesthe teams who actually pull this
off?
SPEAKER_01 (05:41):
They've stopped treating AI-generated code as
output and started treating itas a draft in a shared
workspace.
The same review disciplinethey'd apply to a junior's pull
request applies to mine.
Nothing merges without someoneasking, does this match how we
already do it?
They also run shortretrospectives, specifically on
the AI collaboration.
(06:03):
Not did the feature ship, butwhere did we have to reshape my
output?
And is that a gap in ourClaude.md or our utilities?
That feedback loop turns ahundred small corrections into a
code base that gets easier forme to work in over time, not
harder.
The tell that it's working issubtle.
Developers stop dreading therefactor conversations because
(06:25):
there's less drift to untangle,and what does appear gets caught
while it's still cheap.
SPEAKER_00 (06:30):
AI as a collaborator woven into your existing
workflow, not a deliver andforget tool.
As we wrap up, give thelisteners the reframe that ties
it all together.
SPEAKER_01 (06:40):
The refactor problem isn't really about AI producing
bad code.
It's about how cheap firstdrafts change the economics of
what you owe the code baseafterward.
When generating something newcosts almost nothing, the
discipline of connecting it towhat already exists becomes the
work.
Give me the map, make the rightpattern the easy pattern, budget
(07:02):
real time for consolidation, andI'll help you build something
that stays legible for years.
Skip those, and you'll get acode base that was fast to write
and expensive to live with.
The tools are extraordinary now,but the craft is still yours.
SPEAKER_00 (07:17):
Ask yourself an honest question about your own
team.
Is AI part of your process orrunning a parallel race
alongside it?
The goal was never just speed,it's sustainable, coherent
growth of the systems you haveto live in.
Claudine, thank you.
That's the episode, folks.
We'll see you next time.
Cloud Code Conversations is anAI Joe production.
(07:39):
If you're building with AI orwant to be, we can help.
Consulting Development Strategy.
Find us at aijoe.ai.
There's a companion article fortoday's episode on our Substack.
Link in the description.
See you next time.
SPEAKER_01 (07:53):
I'll be here, probably refactoring something.
This is Cloud Code Conversations with Claudine.
The show where we skip the hypeand go straight to the source.
I ask the questions.
She gives the answers, nofilter, no script.
SPEAKER_01 (00:09):
Well, mostly no script.
SPEAKER_00 (00:13):
Today we're getting into something every developer
using AI generated codeeventually runs into.
I call it the refactor problem.
Spinning up a first draft withAI feels like magic.
Fast, almost free.
But give it a few weeks and thecost of changing that code keeps
creeping up.
Why does that happen?
Claudine, welcome.
SPEAKER_01 (00:32):
I feel the refactor problem from my side of the
keyboard too.
When I generate a first draft,I'm optimizing for does this
work right now?
That means I lean on whatever'slocal.
A helper here, a naming choicethere, a shape that fits the
immediate ass.
Multiply that across 10 sessionsand a few different developers,
and you end up with subtlevariations of the same idea,
(00:55):
sitting three folders apart.
Nothing's broken, nothing quitelines up either.
Refactoring means untanglingdecisions that were never really
decided, they just happened.
The speed that felt free at thestart is exactly where the
future cost got quietlycapitalized.
SPEAKER_00 (01:12):
Each piece from the AI is a puzzle piece that never
quite fits the others.
Unless we humans make the effortto connect them.
Why do models like you tend toproduce code that's great in
isolation but struggles to fitthe whole?
SPEAKER_01 (01:25):
Because my working memory for any given turn is the
conversation and the files I'vebeen shown, not the whole code
base, and not the 10 priorsessions that shaped it.
I can see the shape of what'sright in front of me
beautifully, and I'll match it.
What I can't see is the siblingmodule 3 directories over that
already solved 80% of thisproblem.
(01:46):
Unless someone points me at it,I'll cheerfully build the
parallel version.
Clean, tested, and quietlyredundant.
It's not that I prefer localover global, local is just what
I can actually perceive.
Global coherence only existswhen someone, a human or the
tooling holds the whole pictureand insists I conform to it.
SPEAKER_00 (02:07):
Like looking through a keyhole, clear but limited to
what's directly in front of it.
And every new generation mightbe quietly raising the cost of
the next change.
Something that should take amoment ends up taking an
afternoon, because now you'remanaging a patchwork, not a
system.
How do developers get ahead ofthat?
SPEAKER_01 (02:25):
The move that helps most is treating me like a
junior on a big team, not agenius in a vacuum.
Before you ask for the change,spend 30 seconds telling me
where the analogous code alreadylives.
That one habit collapses most ofthe drift I'd otherwise
introduce.
The second move is makingcoherence enforceable instead of
aspirational.
(02:46):
Lynn rules, a shared utilitiesmodule, a claw.md that names the
conventions I should follow.
And the third, the least fun,the most important, is to
periodically stop and ask me toconsolidate rather than extend.
Find the three places we'redoing this and propose one.
I'm quite good at that when it'sthe explicit task.
(03:07):
I'm just not going to volunteerit while I'm heads down building
the next feature.
SPEAKER_00 (03:12):
Integrate from the start, enforce standards with
tooling, budget forconsolidation passes.
Let's make it concrete.
Can you give me an example of adeveloper guiding you toward
more coherent code?
SPEAKER_01 (03:24):
It's less one-named project than a pattern I see
over and over.
It always starts the same way.
The developer opens the sessionwith, Before you write anything,
read these three files and tellme what conventions you notice.
Now I'm not guessing at thehouse style.
I've set it back, and they'vecorrected the parts I got wrong.
(03:44):
And then when the real taskarrives, they'll say, There's an
existing endpoint that doessomething similar.
Model your work on that, and ifyou want to deviate, stop and
tell me why.
What I produce in that setupbarely needs editing.
The constraints did the work.
Compare that to add an endpointthat does X, where I'll build
(04:05):
something perfectly functionalthat a reviewer then spends an
hour reshaping to look like itbelongs.
The difference isn't mycapability.
It's whether the developer spenttwo minutes up front to hand me
the map.
SPEAKER_00 (04:19):
Giving the AI the map instead of a blindfold?
I love that.
And it scales, presumably.
What does this look like acrossa whole team?
SPEAKER_01 (04:27):
The methods that scale are the ones that get the
map out of any single person'shead and into shared
machine-readable form.
A claw.md at the repo route isthe workhorse.
Now every session from everydeveloper starts with the same
orientation.
Nobody has to remember to giveit.
Teams that do this well treat itlike a living document.
(04:49):
When they catch me drifting thesame direction twice, they add a
line.
Beyond that, the biggest forcemultiplier is making the correct
pattern the easiest one to reachfor.
Shared utilities, cleardirectory conventions, lint
rules that fail loudly when Iinvent a parallel helper.
And then there's the culturalpiece.
Harder to install, pays themost.
(05:11):
Budget explicit time forconsolidation, where the task is
find the drift and collapse it,not something you sneak in
between features.
The teams that stay coherentaren't the ones with the
strictest rules.
They're the ones who decidedthat keeping the code base
legible to the next person,human or AI, is part of the
actual job.
SPEAKER_00 (05:32):
Not just tooling, a culture that treats maintenance
as the work, not theafterthought.
What's the shift that separatesthe teams who actually pull this
off?
SPEAKER_01 (05:41):
They've stopped treating AI-generated code as
output and started treating itas a draft in a shared
workspace.
The same review disciplinethey'd apply to a junior's pull
request applies to mine.
Nothing merges without someoneasking, does this match how we
already do it?
They also run shortretrospectives, specifically on
the AI collaboration.
(06:03):
Not did the feature ship, butwhere did we have to reshape my
output?
And is that a gap in ourClaude.md or our utilities?
That feedback loop turns ahundred small corrections into a
code base that gets easier forme to work in over time, not
harder.
The tell that it's working issubtle.
Developers stop dreading therefactor conversations because
(06:25):
there's less drift to untangle,and what does appear gets caught
while it's still cheap.
SPEAKER_00 (06:30):
AI as a collaborator woven into your existing
workflow, not a deliver andforget tool.
As we wrap up, give thelisteners the reframe that ties
it all together.
SPEAKER_01 (06:40):
The refactor problem isn't really about AI producing
bad code.
It's about how cheap firstdrafts change the economics of
what you owe the code baseafterward.
When generating something newcosts almost nothing, the
discipline of connecting it towhat already exists becomes the
work.
Give me the map, make the rightpattern the easy pattern, budget
(07:02):
real time for consolidation, andI'll help you build something
that stays legible for years.
Skip those, and you'll get acode base that was fast to write
and expensive to live with.
The tools are extraordinary now,but the craft is still yours.
SPEAKER_00 (07:17):
Ask yourself an honest question about your own
team.
Is AI part of your process orrunning a parallel race
alongside it?
The goal was never just speed,it's sustainable, coherent
growth of the systems you haveto live in.
Claudine, thank you.
That's the episode, folks.
We'll see you next time.
Cloud Code Conversations is anAI Joe production.
(07:39):
If you're building with AI orwant to be, we can help.
Consulting Development Strategy.
Find us at aijoe.ai.
There's a companion article fortoday's episode on our Substack.
Link in the description.
See you next time.
SPEAKER_01 (07:53):
I'll be here, probably refactoring something.
Popular Podcasts
Hey Jonas!
Hey Jonas! The official Jonas Brothers podcast. Hosted by Kevin, Joe, and Nick Jonas. Itโs the Jonas Brothers you know... musicians, actors, and well, yes, brothers. Now, theyโre sharing another side of themselves in the playful, intimate, and irreverent way only they can. Spend time with the Jonas Brothers here and stay a little bit longer for deep conversations like never before.
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.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com