Add the next one
You have one loop running. This is the order to add the rest, so you never face thirty five at once.
Launch one, prove it, add a second, then fan out. Five engineers each quietly building their own docs loop is five bills and five sets of rules drifting apart.
The next two, same chapter, same runner
Both are starred in the field guide for the same reason loop 2 is: mechanical, fully checkable, trivially reversible.
| Next | Loop | Why it is safe |
|---|---|---|
| 2nd | 6 lint, format, and type fixes | Pure busywork. The tools decide what is right, not the agent. |
| 3rd | 7 release notes and changelog | It writes prose from your commit log. Nothing executable changes. |
One command each, and the runner and the secret are already there:
./install.sh ~/code/my-app 06
./install.sh ~/code/my-app 07
Then ./run-loop.sh 06 --check in your repo. Run it again with more loop numbers whenever you
want another; it never overwrites your work without asking, and it adds only the settings the
new loops need to your existing loops.env.
Then adopt a whole chapter
Once three loops feel boring, take a chapter. Pick the one that matches the pain you actually have.
| Chapter | Loops | Adopt it when | Start here |
|---|---|---|---|
| 1 · Software engineering | 1-9 | your maintenance backlog never shrinks | chapter |
| 2 · Cloud and platform | 10-18 | your cloud bill or your on call is the problem | SETUP.md first |
| 3 · AI and ML engineering | 19-26 | you ship prompts or models and have no evals | chapter |
| 4 · QA and testing | 27-35 | your CI signal is not trusted any more | chapter |
Each chapter has its own README.md (how the pack works) and LOOPS.md (one row per loop, with
which ones ship working code today).
Cloud is the one that is different
Cloud loops act on your AWS account, not a git repo. They sign in with a scoped OIDC role rather than an API key in a secret, and most of them are Flags, you decide, because most cloud actions cannot be taken back.
Do loop-packs/cloud-loops/SETUP.md first, and try everything with DRY_RUN=1 before you point
anything at a live account.
Before you build one of your own
The catalog is thirty five loops, not a limit. When you want to add your own, run it through the guide's five question test (page 19) first:
- Does it recur, and does it actually hurt?
- Can it check its own work, cheaply and honestly?
- Can you undo a miss?
- Is a boring deterministic tool already doing it?
- Bonus: does the output compound?
One "no" on the first four and it is not a loop. It is a chore in a loop costume, and building it burns money you will not get back.
Then do the napkin: cost is tokens per run times runs per month. Payoff is the hours removed plus the failures prevented. If the payoff does not clear the cost by a wide margin, do not build it. More on that in operating a fleet.
← Your first loop · Contents · Next: Operating a fleet →
The 35 loops: Engineering · Cloud · AI and ML · QA
Reading is the slow way round. This runs all four chapters against seeded data:
git clone https://github.com/fakiolinho/From-Prompt-to-Loop.git && cd From-Prompt-to-Loop && ./run-all-demos.sh