steelmorgan/skill-learning-policy

Knowledge Accumulation Policy (Skill Learning)

TRIGGER — WRITE a lesson (MUST, strict)

Completed a work cycle with ≥2 iterations (wrote → error → found a correct path → success) → retrospective is MANDATORY. Apply the skill-learning skill (framework/skills/agent-process/skill-learning/SKILL.md) and carry it out.

The task was solved on the first try → retrospective is NOT needed, the trigger does not fire.

Trigger acceptance: in full-cycle, skipping the retrospective is recorded by the orchestrator during phase acceptance. In quick-fix/FREE there is no orchestrator acceptance → the agent confirms completion itself with a line in {role}-context.md / report: SKILL_LEARNING: matched|refined|new|skipped(<reason>).

TRIGGER — READ lessons (before working with a skill)

Before starting work with any skill, read the accumulated lessons of the skill owner, BEFORE design/implementation:

• <skill>/references/learned-patterns.md — general techniques;
• {project}/.context/learned-patterns.md — project techniques.

Treat confirmed as additional rules, candidate as hints. If the file does not exist, skip it. This is a phase trigger: on entering work with a skill, not on every turn.

MUST (invariant, always)

• Abstraction level above the incident. A lesson describes a named CLASS of error + an anti-recipe (generalized prohibition/check), NOT a specific incident. Concrete file:line / symbol names from that incident go as a "source example", NOT as the body of the rule.
• Abstraction test (record acceptance). The anti-recipe must apply to at least ≥1 OTHER hypothetical situation besides the one that generated it. If not applicable, this is an instance: reformulate it at a higher level or reject it.
• Reconciliation update before recording. Reconcile the incident with existing recipes (references/learned-patterns.md of the skill owner + {project}/.context/learned-patterns.md). If it falls under an existing class, update it (candidate→confirmed, expand the anti-recipe boundaries, add the source), do NOT create a duplicate. If it does not fit, create a new class. The outcome must be recorded explicitly: matched | refined | new.
• Records are written ONLY to references/learned-patterns.md of the skill owner (general) or {project}/.context/learned-patterns.md (project) — NOT into the body of SKILL.md.
• Target directory for a general lesson is the RU source of the framework (the skill owner's directory in framework/, NOT the installed ENG symlink framework_eng/: otherwise the lesson is lost during synchronization). Search for the directory and synchronize into ENG using the skill-editing-from-project skill. The project lesson ({project}/.context/learned-patterns.md) is in Russian, without mapping or synchronization.
• Acceptance and anti-acceptance are one record; the first case is candidate, repetition is confirmed.

The retrospective procedure (reconstructing the iteration chain, abstracting to a class, abstraction test, reconciliation cascade, choosing the level and skill owner, searching the RU directory) is in the skill-learning skill.

depends_on:

• framework/skills/agent-process/skill-learning/SKILL.md
• framework/skills/framework-meta/skill-editing-from-project/SKILL.md