| name | steering-correction-promotion |
| description | On user steering corrections or remember-this, pick the smallest durable artifact (skill section, scoped rule, or rare always-apply rule), slim AGENTS.md, optimize agent-brain context, then prepend JRN. Use instead of creating new alwaysApply rules by default. |
| disable-model-invocation | true |
Steering correction promotion
Run this skill in the same turn as any user steering correction — how you decide, operate, route, log, or commit. Finish with artifact + JRN + brief confirmation.
Also invoked from agent-self-improvement.
Context budget (optimize every promotion)
| Layer | Cost | Use for |
|---|
alwaysApply: true rules | Every turn — add sparingly; merge or scope when possible | Routing, portable paths, Python env — only what every task needs |
globs rules | When matching files are in play | Script execution, docs schema, test patterns |
| Skills | When task matches or agent opens skill | Procedures, WSL GPU, promotion workflow itself |
AGENTS.md | Router only — stay thin | One index table; no procedure prose |
| Self-journal | Receipt — not loaded every turn | Audit trail with Artifact path |
Default: prefer skill section or scoped rule over a new alwaysApply rule.
Decision tree
Answer in order; stop at the first fit.
1. Multi-step procedure (train, debug playbook, log EXP)?
→ Extend existing skill OR new skill + skills README row
→ AGENTS.md: new index row ONLY if nothing routes there yet
2. Only matters when editing/running certain paths?
→ Scoped rule (.mdc with globs, alwaysApply: false)
→ Merge into existing scoped rule if same glob family (e.g. scripts/**)
3. Universal but <5 lines and same topic as an existing alwaysApply rule?
→ Merge into that rule — do NOT add a new file
4. Truly required on EVERY turn (safety, entry routing)?
→ alwaysApply rule — LAST RESORT
→ Keep under ~40 lines; regenerate [agent-brain.md](../../../docs/agents/agent-brain.md)
5. One-off preference for this session only?
→ Chat — no JRN unless user says remember this
6. Mechanical bug with a code fix?
→ Code/tests first; optional scoped rule if agents keep repeating it
Same-turn workflow
| Step | Action |
|---|
| 1 | Run decision tree — pick smallest enforcement layer |
| 2 | Implement — edit skill/rule/AGENTS/code; dedupe while editing |
| 3 | Optimize — see checklist below |
| 4 | JRN — prepend with Artifact path(s) |
| 5 | Confirm — artifact + journal id + what was not promoted to alwaysApply |
| 6 | Quick refresh — agent-brain-refresh steps 1–3 |
Optimization checklist (required)
After every promotion, scan agent brain for waste:
AGENTS.md rules
| Do | Don't |
|---|
| Add a row to the task index when a new skill/index is the entry | Paste skill or rule content into AGENTS.md |
| Link agent-brain.md for rule inventory | Duplicate alwaysApply lists (Cursor loads them automatically) |
| Link north star / pre-submit once | Grow AGENTS.md with journal or EXP content |
Rule catalog
Update docs/agents/agent-brain.md by hand during refresh. Verify with read-only python scripts/audit_agent_brain.py. Scoped summary: skills README § Agent brain.
JRN receipt
| **Artifact** | Path(s) — include layer chosen (skill / scoped rule / alwaysApply / code) |
| **Action taken** | One line: what enforces the correction + any demotion/merge done |
Signals
Steering: “remember this”, “don’t …”, “always …”, “stop doing X”, corrections to agent process (not hyperparameters or one-off task args).