# Why I replaced my 20-step AI agent prompt checklist with a 1976 Unix Makefile — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/d5e5d6b7528288e6

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-09 00:50 ingested 2026-10-10 02:34

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1x17saa/why_i_replaced_my_20step_ai_agent_prompt/)

## retained text



```
When I first started building autonomous agent skills for my Dart and Flutter repos, I wrote them the way everyone does: a numbered checklist from Step 1 (pick an issue) to Step 17 (triage review bots and merge). It looked great until real life happened. Whenever a review bot left a comment after Step 9, I had to write awful imperative GOTO instructions ("jump to Step 12.2, go back to Step 5, don't redo Step 1"). Worse, if I closed my laptop and woke the agent up two hours later on an open PR to fix a failing CI check, it would read Step 1 at the top of the prompt and ask me which issue we were working on.

It hit me recently that Stuart Feldman solved this exact problem at Bell Labs in 1976 with `make`. A software workflow isn't a linear checklist—it's a Directed Acyclic Graph of physical artifacts and prerequisites.

When you write the agent's skill as an actual Makefile DAG (`finish` depends on `land`, which depends on `ready-to-land` and `human-landing-approval`, all the way down to `approved-plan` and `ticket-assigned`), two things happen immediately. First, LLMs have seen millions of Makefiles in pre-training, so they resolve `target: prerequisites` way more reliably than prose rules. Second, you get rid of loop counters and step numbers completely. If a bot finds a bug and the agent edits a file, the working tree is now dirty—which physically invalidates the prerequisites for `commit` and `approved-adversarial-report`. The graph forces the agent back through the code critic and linter automatically.

One weird failure mode we did hit along the way: if a target has 5 or more prerequisites on a single line, the model builds up "verification momentum" after checking the first four and blows right past the human approval gate at position 5. Capping every target at 2 to 3 prerequisites (and factoring human gates into strict 2-item `action: ready-state human-approval` pairs) fixed it completely across our last 140+ tickets.

How are you all structuring long multi-step workflows in Claude Code, Cursor, or Antigravity so the agent doesn't lose its place mid-PR?
```



document fetched 2026-10-10 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1x17saa/why_i_replaced_my_20step_ai_agent_prompt
