skills.sh
AvailableUniversal skill installation, including Claude Code and Codex.
npx skills add https://github.com/ulpi-io/skills-autonomous-engineering --skill auto-learnauto-learn
Turns structured run evidence into verified, deduplicated lessons that load automatically next time.
WORKFLOW / 05/auto-learn [checkpoint path or run id to harvest] (default: the most recent .ulpi/runs/*.json)
- 01
Harvest candidates from checkpoints, open items, guard blocks, fix counts, and degradation records.
- 02
Adversarially verify that each candidate is true, general, and actionable.
- 03
Deduplicate, merge, prune superseded guidance, and cap additions at five.
- 04
Route project, area, and environment lessons into native auto-loaded context.
- 05
Surface machine defects to the user and leave the next run a smaller rediscovery burden.
▸ operating envelope
Use it deliberately.
The shortest useful definition of where this skill helps—and where it should stay out of the way.
Use when
- An autonomous run has completed and produced checkpoint evidence.
- Blocks, guard refusals, thrash, or delegation gaps should change a future decision.
- The full pipeline is closing and needs its feed-forward learning pass.
Do not use when
- The run is still in progress and the evidence is incomplete.
- The candidate merely repeats existing repository documentation.
- The source is an intuition, secret, one-off flake, or unsupported anecdote.
▸ output contract
What must be true when it stops.
Produces
- Harvested candidate and evidence ledger
- Verified survivors and rejection reasons
- Merged lessons in native auto-loaded context
- User-facing machine-defect report
Guarantees
- Every candidate carries structured evidence from a completed run.
- False, one-off, or non-actionable candidates are rejected before writing.
- Lessons land in CLAUDE.md, path-scoped rules, or Claude auto memory—not an inert bespoke file.
- Human-written memory is preserved and secrets are never learned.
- Memory remains curated, deduplicated, budgeted, and capped to five additions per run.
Honest failure states
- rejected — the candidate is false, too specific, or not actionable
- duplicate — an existing memory already captures the lesson
- machine_defect — the fix belongs in a skill, guard, or template and is surfaced to the user
- unsafe_content — the candidate contains secrets or would overwrite human-owned memory
▸ enforcement
Know which claims are executable.
Scripts are linked when this skill owns deterministic machinery. The remaining rules are explicit operating contracts and must not be marketed as hooks.
▸ install and invoke
One skill, three honest surfaces.
Claude Code plugin
AvailableInstall the full plugin, then invoke /auto-learn or let routing select it from context.
Codex plugin
Work in progressA work in progress on the codex-native-plugin branch — build it, start a new session, then invoke $autonomous-engineering:auto-learn.
▸ composes with