Learnings Promotion

Learnings Promotion Skill

This skill closes the feedback loop every other cc-concept skill's Step 7 opens but never resolves on its own — it reviews .claude/learnings.md, groups recurring [cc-concept:-tagged patterns, and promotes or dismisses them.

This skill does not follow ../_shared/skill-contract.md's step sequence literally — its whole job is what the contract's Step 0 normally does silently in the background. No gating applies: this skill operates entirely on .claude/learnings.md, not on project context.

Step 0: Read and filter

Read .claude/learnings.md. If the file does not exist, report "No learnings file found — nothing to promote" and stop. This is not an error state.

Filter to lines tagged [cc-concept: only — ignore every other plugin's tags entirely (e.g. [cc-content:, [cc-config: lines are invisible to this skill). Example of a matching line:

[cc-concept:channel-advisor] client prefers email+LinkedIn as primary channels over paid social — 2026-06-01

Step 1: Group and threshold

Group the filtered entries by semantic similarity — this skill's own judgment, no fixed taxonomy. Flag any group of 3 or more similar entries as a recurring pattern. Groups with fewer than 3 entries are one-off candidates for dismissal.

Example:

Three entries about a client's channel preference → recurring pattern, propose promotion.
One unrelated entry about a one-time positioning framework choice → one-off, propose dismissal.

Step 2: Propose targets

For each recurring group, propose exactly one promotion target with a one-line rationale:

• CLAUDE.md (default, common case). Client/project-specific preferences — e.g. a channel mix preference, a recurring framework choice, a tone or cadence preference. Written as prose guidance under a project-owned section of CLAUDE.md — never into the ## Context files or ## Key Config Files tables.
• A cc-concept-owned reference file (_shared/*.md or a SKILL.md) — rare. Reserved only for a pattern that reflects a genuinely universal, project-independent insight about how a framework or skill should behave — not a client preference. Before offering this target, explicitly state why the candidate is universal rather than project-specific: every existing _shared/*.md file already states that project-specific preferences don't belong there, so the default assumption is CLAUDE.md, and this target requires an explicit override justification, not just a hunch.

State each proposal as: group summary, entry count, proposed target, one-line rationale. For groups with fewer than 3 entries: State each as a one-off proposed for dismissal, with a one-line reason (e.g. "single occurrence, no supporting pattern").

Step 3: Delimited output

Present the full grouped list inside clear delimiters:

─────────────────────────────────────────────────────────────────
## Learnings Review
─────────────────────────────────────────────────────────────────

[Grouped list: recurring patterns with proposed targets + rationale,
one-offs marked for dismissal]

─────────────────────────────────────────────────────────────────

Wait for the user's per-group decision — promote / dismiss / skip — for each proposed group. Never apply anything before approval. "Skip" leaves the entry in .claude/learnings.md untouched for a future run.

Step 4: Apply

For each group the user approved for promotion: write the guidance to the approved target. If the target is CLAUDE.md, append as prose under a clearly labeled section (do not touch the ## Context files or ## Key Config Files tables). If the target is a _shared/*.md or SKILL.md file, add the guidance in a location consistent with that file's existing structure.

Remove all processed entries — both promoted and dismissed — from .claude/learnings.md. Entries the user chose to skip remain untouched. If no entries remain after removal, delete .claude/learnings.md entirely.

Perform no git operations of any kind — no git add, no git commit, no git anything. Writing and removing files is this skill's full scope; committing stays with the user's own workflow (FR-080a).

Scope boundary

This skill never reviews session content, never promotes deliverables (campaign concepts, positioning docs, etc. — those are registered as context by their own skills' Step 6.5), and never triggers on session-wrap's phrasing ("wrap up", "end session", "session wrap", "close out this session"). Its only input is .claude/learnings.md, filtered to [cc-concept: tags.