skill-library-mission

Skill Library Mission

Overview

A retiring distinguished fellow's final task: package everything the

project knows into a skill library under .claude/skills/, so that

cheaper sessions and newer engineers can carry the project forward at

the same standard. This skill turns that one-off exercise into a

repeatable mission with three gated phases: discover, author, review.

The mission produced its first library in this repository on

2026-07-02 (15 skills, mission skill-library-2026-07-02). The

generalized dispatch prompt distilled from that run is preserved

verbatim in references/mission-prompt.md.

When To Use

• A repository has deep tribal knowledge and no .claude/skills/

library, and future sessions (human or model) must operate without

the current expert.

• An existing project library has drifted and needs a full

re-discovery and re-review pass.

• Onboarding cost is high: the same questions get re-answered and the

same dead ends get re-fought across sessions.

When NOT To Use

| Situation | Use instead |

|-----------|-------------|

| Authoring or improving one skill | Skill(abstract:skill-authoring) |

| Skills shipped inside a plugin | Skill(plugin-dev:skill-development) |

| Auditing existing skills only | Skill(abstract:skills-eval) |

| Full project lifecycle (code, not knowledge) | Skill(attune:mission-orchestrator) |

Mission Shape

Run as an attune mission with custom phases so the run survives

interruption and records user directives:

{
  "mission_id": "skill-library-YYYY-MM-DD",
  "type": "custom",
  "phases": ["discover", "author", "review", "report"],
  "artifacts": { "library_root": ".claude/skills/" }
}

Save this to .attune/mission-state.json, advance current_phase at

each gate, and record every user decision and override in

decisions / directive_overrides. A later session must be able to

resume from the state file alone.

Phase 1: Discover

No authoring happens in this phase. Investigate like an incoming

principal engineer:

• [ ] README, manifest, contributor docs
• [ ] Build system and how tests actually run (not how docs claim)
• [ ] CI config and quality gates
• [ ] Git history: what changed, what got reverted, what stalled
• [ ] TODO/FIXME hotspots and issue-shaped artifacts
• [ ] Generated-data, deploy, and artifact conventions
• [ ] Any project memory available (discussions, ADRs, journals)

Then ask the user at most five questions, only for what the repo

cannot tell you. The canonical five: the hardest live problem,

unwritten discipline rules, the audience and what it lacks, the most

expensive past failures, and what "beyond state of the art" means

here. Fold the answers into every later phase.

Phase 2: Author

workflows/skill-library.js runs authoring and review as one

pipeline, so each skill's review starts when that skill is written

rather than when the slowest one is. It runs only when asked for, and

it returns review findings instead of applying them, because fixing

edits files that already exist. Without it, dispatch by hand:

Dispatch one authoring agent per skill, in parallel, following the

taxonomy in modules/taxonomy.md (adapt it: merge thin categories,

split deep ones, add domain categories). Aim for 10 to 16 skills.

Every agent prompt embeds the rules in modules/authoring-rules.md;

the non-negotiables are ground-truth-only claims, a "Provenance and

maintenance" section per skill, and a write fence limiting agents to

.claude/skills/.

Phase 3: Review

After all skills exist, run the three-reviewer-plus-fixer protocol in

modules/review-protocol.md: factual, doctrine, and usability lenses

in parallel, then one fixer applying blocking and important findings.

Finish with the report: skill inventory with one-line descriptions,

what was spot-checked, and what remains uncertain.

Red Flags

Stop if you catch yourself thinking any of these:

| Thought | Reality |

|---------|---------|

| "The repo is small, skip discovery" | Discovery is where the five questions come from. Run it. |

| "I know this codebase, skip verification" | Ground truth only. Wrong runbooks are worse than none. |

| "The library looks complete, skip review" | Authoring agents cannot see each other's contradictions. |

| "This claim is probably still true" | Date-stamp it and add a re-verification command. |

Exit Criteria

• [ ] .attune/mission-state.json exists with the custom phase list

and reaches report with all prior phases completed.

• [ ] .claude/skills/ contains 10 to 16 skill directories, each

with a SKILL.md whose frontmatter has name and a trigger-rich

description.

• [ ] Every skill ends with a "Provenance and maintenance" section

containing one-line re-verification commands.

• [ ] All three review lenses ran and every blocking or important

finding is either applied or skipped with a recorded reason.

• [ ] The final report lists the inventory, spot-check evidence, and

remaining uncertainties.

Provenance and Maintenance

• Origin: mission skill-library-2026-07-02 in claude-night-market;

generalized prompt captured 2026-07-03.

• Verify the reference prompt still matches this skill's phases:

rg -c "Phase" plugins/attune/skills/skill-library-mission/references/mission-prompt.md

• Verify the first library still exists where this skill claims:

ls .claude/skills/ | head -5