SuperAPI AI Directory

lesson-propose

Agents & models

View repository

Updated 10-16 06:17

About this skill

Lesson Propose

The worker proposes lessons; it never promotes them. Promotion (candidate → control/approved_lessons/ or rules.md) is a human git action, not a tool the worker can call.

When to use

  • Task wrap-up, postmortem, or when a correction/drift note should outlive this task.

Procedure

  1. Read state/tasks/<task_id>/correction_journal.md and any session observations.
  2. For each keeper, write a candidate into state/tasks/<task_id>/candidate_lessons.md, tagged with a target tier:
    • working-only — relevant to this task only; stays in state/.
    • approved-lesson-candidate — recurring; propose into control/approved_lessons/.
    • policy-candidate — stable, cross-task; propose into control/rules.md.
    • gate-candidate — binary, machine-checkable; a human may translate it into an agent-completion-gate check (this skill does NOT touch the gate).
  3. Each candidate carries: the lesson, evidence (cite the journal entry), applies_when, and not_authoritative_until_reviewed: true.
  4. Leave everything in candidate_lessons.md (pending). Do not move anything into control/ yourself.

Outputs

  • candidate_lessons.md with tier-tagged proposals + evidence.
  • A short summary the human can review and promote in git.

Do not

  • Do not write to control/approved_lessons/ or control/rules.md — read-only to the worker; promotion is a reviewed action.
  • Do not edit any agent-completion-gate manifest. Surfacing a gate-candidate is a suggestion to a human, not an action.
  • Do not mark a lesson "enforced" without a corresponding rule/gate actually carrying it.