Création et gestion de compétences expertes

---

description: create, refine, review, or retire reusable skills at expert level allowed-tools:

• "mcp__pk-qmd__*"
• "mcp__obsidian__read_note"
• "mcp__obsidian__write_note"
• "mcp__obsidian__search_notes"
• "mcp__obsidian__manage_tags"
• "mcp__llm-wiki-skills__*"
• "mcp__brv__*"

---

Read CLAUDE.md, then read LLM_WIKI_MEMORY.md if present, SKILL_CREATION_AT_EXPERT_LEVEL.md if present, .llm-wiki/config.json, wiki/index.md, wiki/skills/index.md, and recent wiki/log.md.

Retrieval Protocol (apply before every search)

Step 0 — Context gate: Check wiki/skills/index.md and the current conversation for a matching skill before any tool call.

Step 1 — Cheap tier: If the skill MCP server is available, call skill_lookup first. If it resolves the question fully, stop — do not also fire pk-qmd.

Step 2 — Standard tier: Only escalate to pk-qmd if skill_lookup misses or the skill MCP server is unavailable. One pk-qmd lex call to find related skill pages or evidence.

Step 3 — Full tier: pk-qmd lex + vec for novel skills with no prior art. Cap at 2 hops. brv in parallel only if user workflow preferences affect the skill shape.

Create or update a reusable skill from the current task, trajectory, or evidence:

• if the local skill MCP server is available, call skill_lookup before exploring
• for long tasks or expensive exploration, call skill_reflect or skill_pipeline_run first so the important context is captured as a reducer packet plus artifact refs
• call skill_validate before direct save when the candidate is non-trivial, likely duplicated, or needs explicit review
• use skill_propose, skill_feedback, or skill_retire for lifecycle operations
• search for existing related skills first
• write or update the skill page under wiki/skills/active/
• append reasoned review notes under wiki/skills/feedback/ when the task is feedback-driven
• retire the skill into wiki/skills/retired/ when evidence shows it is unsafe or the score should fall below the retirement threshold
• update wiki/skills/index.md
• append a skill entry to wiki/log.md

Skill requirements:

• optimize for learn-once, reuse-forever shortcuts
• include trigger, preconditions, fast path, failure modes, and evidence
• run a privacy gate before saving
• validate duplicates before saving, and merge deltas when overlap is strong
• for long tasks, capture a strong middle-manager reducer packet with an explicit route_decision
• prefer a 1-3 call reusable recipe over verbose narrative
• mark HTTP upgrade candidates explicitly, but do not claim them without evidence

Post-save Sync

After creating or editing any skill page under wiki/skills/active/, determine whether the skill belongs in the canonical pk-skills1 source of truth (C:\Users\prest\.agents\skills1\pk-skills1).

Decision rule:

• Local operational shortcut for this vault only → no sync needed. Note in the skill's review trail.
• Reusable across sessions and projects → promote to pk-skills1, then launch skill-sync-manager as a background sub-agent to propagate to all mirrors.

When syncing, launch skill-sync-manager with this brief (fill in concrete paths):

Source of truth: C:\Users\prest\.agents\skills1\pk-skills1
Changed skills:
- <skill-root>

Destinations to check:
- C:\Users\prest\.codex\skills
- C:\Users\prest\.pi\agent\skills\pk-skills1-imported
- C:\Users\prest\.agents\skills
- C:\Users\prest\.claude\skills

Backup location: C:\dev\Desktop-Projects\Helpful-Docs-Prompts\skills1-backup

Constraints:
- validate each changed skill before syncing
- safe add/update only — never delete without explicit approval
- preserve destination-local customizations
- create a dated zip backup entry before mirroring

Return: changed skills, validation results, actions taken, approval-required follow-ups

Divergence rule: Update the source-of-truth package first. Combine any mirror-only content into pk-skills1 before mirroring outward. Never edit mirrors first.

If background sub-agents are unavailable: run managed-skill-sync inline:

python .codex\skills\managed-skill-sync\scripts\audit_and_sync.py --json-out .artifacts\skill-sync-audit.json --apply

For the current session's adaptive-retrieval-routing skill: Ask the user whether to promote it to pk-skills1 before syncing.

Routing

• use pk-qmd MCP tools for repo-local evidence retrieval and prior skill lookup
• use pk-qmd first when the right prompt, note, file, or skill page is not known yet
• use obsidian MCP tools for vault reads and writes when available; fall back to direct file I/O if obsidian is down
• use GitVizz when repo topology or API context sharpens the reusable recipe
• use brv only for durable user or workflow preferences that materially affect the skill
• if pk-qmd and brv conflict, trust current source evidence

Return:

• stack/config used
• MCP tools used (or fallback note if obsidian was unavailable)
• files read
• files changed
• what changed
• unresolved questions