Worked example — the How-To Guide Library
A runnable instance of the 11ai-automation method in a simple domain:
an agent that grows a library of how-to guides (each guide teaches one task,
e.g. "set up a systemd user timer"), where every command in every guide has
actually been executed before it ships.
Read the two files in examples/ side by side with the core
skill's conductor template — every <…> blank is filled in:
- examples/CONDUCTOR.md — the complete spec.
- examples/ledger.json — a ledger mid-flight: one guide done, one mid-build, config with caps.
- examples/runner.conf — the provider config: one
AGENT_CMDline makes the same automation run on any headless agent CLI, with an optional second provider as the reviewer.
How the pattern maps to this domain
| Concept | Here |
|---|---|
| Work item | One how-to guide (guides/<id>/guide.md) |
| Milestones | 1: core walkthrough (tested commands), 2: failure modes + troubleshooting, 3: variants (other OS/platform) |
| Depth bar | Every command block executed with output captured to guides/<id>/transcripts/; every link resolves; a stranger can follow the guide start to finish on a clean machine |
| Quality gate | markdownlint, a link checker, and node scripts/verify-guide.mjs <id> (asserts structure, transcript coverage, no placeholders) |
| Review | One fresh-context reviewer per tick; must re-run the verifier and re-execute at least one command block itself |
| Runner | run-tick.sh with AGENT_CMD from runner.conf — works with any headless agent CLI |
Using it as a starting point
- Copy
examples/into a fresh repo; rename the mission, theme, and paths. - Swap the depth bar and gate commands for your domain (that is usually the only real design work).
- Prove one tick by hand, then schedule per the core skill's runner reference.
The fastest way to design a new automation is to diff your domain against this one: what is the item, what are its milestones, what would a stranger need to be able to do, and what command proves it.