Mike Litman
The Compounding Work
Mike Litman · 2026

The Compounding
Work

The work that shapes every session afterwards

The visible work

The site is live.
The deck is deployed.
You shipped.

This is what progress looks like. Fast, visible, satisfying. AI made it possible to ship things in hours that used to take weeks.

Buggy Smart Queue Index First Order London With Moshi 25 deployed sites Built with Claude Code
The invisible work

Nobody sees the hooks.
Nobody sees the audits.
Nobody sees the rules that held.

That is exactly the point.

The entry point

Two screenshots.
One contradiction.
One question worth asking.

A billing email said 100% of GitHub Actions minutes were used. The budget page showed $0.18 spent of a $100 budget. Both were correct. They were measuring different things. And the gap between them contained three weeks of invisible waste.

May 2026 3,000/3,000 minutes used 20 days early

What the question found

Before: 74 minutes
of compute.

37 full pipeline runs in one day
2 minutes per deploy: bake, sanity check, deploy, smoke test
3,000 GitHub Actions minutes exhausted 20 days early
Nobody watching the monthly total

After the workflow fix

18 minutes.
Same output.

37 lightweight deploys
30 seconds per deploy
85% reduction in compute
Two workflow files changed

74 min ÷ 18 min, from gh run list audit, May 2026

The structural gap

AI optimises locally.
You optimise globally.
Nobody was doing both.

I had deployed the same site 37 times in one day without noticing. The session goal was: get the deploy done. It got done every time. The cumulative cost of how it got done was invisible until a billing email looked wrong and I followed the question.

The same pattern, three times

It is not just compute.

Compute
37 redundant workflow runs
On one shipping day. The AI was optimising locally. Nobody was watching the monthly total.
Architecture
Railway: wrong repo, 5 weeks
GitHub Actions deployed cleanly. The site looked fine. The wrong code was running the whole time.
01
Case study

Five weeks.
Wrong repo.
No errors.

Buggy Smart's Railway service was pointed at an abandoned repository while all development continued in the active one. GitHub Actions deployed cleanly. The site looked live. The wrong code ran every day. The AI session had no way to see this: it was optimising within the session, not auditing the full deploy chain.

The fix was not to check more carefully next time. The fix was to make "which repo?" a verified fact, not an assumption.

April 2026 Cost spike: ~260 GBP/month Root cause: wrong source repo
02
Case study

The rule that
kept breaking.

A deck was built without running the mandatory 4-round quality process. The process was written in CLAUDE.md. It was read at the start of the session. It was ignored during the build. When the rule was pointed out, the answer was not to add more text. The answer was to make the rule un-skippable.

May 2026 Root cause: instruction-only enforcement Fix: hook + marker system
The decision

The answer was not
more text in CLAUDE.md.
The answer was a hook.

Every time a rule broke, the instinct was to rewrite the instruction more clearly. Clearer instructions failed just as reliably. The shift was recognising that the problem was the medium, not the message. Instructions live in text. Infrastructure lives in code.

"Add more text to the instructions, see if there is a stronger, more robust approach and build a hook."

Mike Litman, 2026-05-11

The principle

Every rule that requires
memory will fail.

Every rule the harness enforces will hold.

Stop trying to remember. Start designing systems where remembering is not required.

Instructions as Infrastructure in production

26 live hooks

Each hook is a rule that cannot be broken. Not by accident. Not by a session that got busy. Not by an AI optimising locally.

  • 01No deck ships without 4 rounds of quality review + personal verdict (built after a deck shipped 80% complete)
  • 02No blog post ships with a fabricated biographical timeframe ("eighteen months" -- invented, published, then caught)
  • 03No credential ever reaches a commit (after a real token leak in April 2026)
  • 04No workflow file is overwritten without reading git log first (a prior session replaced a hardened pipeline)
  • 05No em dash ever appears in published prose (broke in every session until the hook made it impossible)
26
Live hooks in production
Zero
Memory-dependent rules
Five
Critical incidents caught
Every
Session, automatically
The objection

“Engineering teams have always done this.”

Yes. With code review. Pull request gates. Colleagues who catch mistakes. Senior engineers who hold the institutional memory.

This is that, for a team of one working with a system that has no memory between sessions and optimises locally by design.

The hooks are not a productivity trick. They are the institutional memory the AI cannot hold.

The audit as practice

Confusion is a signal.
Follow it.

Start here tomorrow: open your GitHub billing page. Check Actions minutes used this month. If it is over 50% before mid-cycle, run the workflow audit below. That is your entry point.

  • 01GitHub Actions: sum minutes per workflow, not per repo.
  • 02Deployment services: verify the source repo is the active one.
  • 03Netlify: check for redundant deploys, not just 200 responses.
  • 04Scheduled jobs: read the actual log output, not the success tick.
  • 05Cadence: once per billing cycle. Trigger: any moment of confusion.
The compounding effect

Compounding means:
the work pays interest.

Day 1
You build a hook
30 minutes. One rule enforced. One incident prevented, eventually.
Month 1
The hook fires silently
Dozens of sessions. Each time, the rule held without you thinking about it.
Month 6
26 hooks. Every rule enforced.
The floor is solid. Infrastructure maintains standards, not you.

What the human's job becomes

Execute
tasks.

Write the code. Run the deploy. Check the output. Repeat. The work was in the doing.

Now

Design
conditions.

Build the hooks. Run the audits. Ask the confusing question. The work is in making the doing better, permanently.

High agency applied to the system, not just the output
The constraint is no longer time; it is judgment
The Compounding Work

The floor under output
is rarely the headline.
Build the floor.

hello@mikelitman.me · mikelitman.me

Continue from slide ?