phrazzld/curate

/curate

Evolve the Spellbook library. This is the primary workflow for maintaining and growing the primitive collection.

Routing

| Command | Action | |---------|--------| | /curate | Full curation session (all phases below) | | /curate scan | Scan external sources for new skills worth indexing | | /curate improve | Synthesize observations into discrete improvements | | /curate brainstorm | Identify gaps, propose new primitives | | /curate audit | Review existing skills for quality, overlap, staleness | | /curate research | Research power users, best practices, ecosystem trends |

Full Curation Session

Phase 1: Ecosystem Scan

Research what's new in the agent skills ecosystem:

1. Check indexed sources — Run python3 scripts/generate-embeddings.py --dry-run to see current source coverage. Are there new skills in existing sources?

2. Discover new sources — /research web-search for:

   • New skill repos on GitHub (search "SKILL.md" agent skills)
   • skills.sh marketplace for trending skills
   • Posts/threads about effective Claude Code / Codex workflows
   • Power user setups and custom agent configurations

3. Evaluate candidates — For each discovered skill/source:

   • Does it fill a gap in our library?
   • Is it high quality? (check frontmatter, structure, references)
   • Does it overlap with something we already have?
   • Would our users actually use it?

4. Update sources — If a new repo is worth indexing, add it to EXTERNAL_SOURCES in scripts/generate-embeddings.py and regenerate.

Phase 2: Observation Synthesis

Process accumulated feedback from consuming projects:

1. Collect — Find all .spellbook/observations.ndjson files:

   find ~/Development -name "observations.ndjson" -path "*/.spellbook/*" 2>/dev/null
   

2. Cluster — Group by primitive FQN. Primitives with 2+ observations have a real pattern.

3. Act — For each cluster:

   • High confidence (>= 0.8): Direct PR with the fix
   • Medium (0.5–0.8): Create a GitHub issue with evidence
   • Low (< 0.5): Keep logging, note the pattern

Phase 3: Brainstorm

Identify gaps and propose new primitives:

1. Run /research thinktank with:

   • Current skill inventory (from index.yaml)
   • Recent user workflows and pain points
   • Question: "What recurring workflows are NOT covered by a skill?"

2. Cross-reference with ecosystem — Compare our coverage against anthropics/skills, openai/skills, vercel-labs/agent-skills. What do they have that we don't? What do we have that's unique?

3. Propose — For each candidate new primitive:

   • What gap does it fill?
   • Who would use it? How often?
   • Could an existing skill be extended instead?
   • Effort estimate

Phase 4: Audit Existing Skills

Review the current library for quality and relevance:

1. Overlap detection — Run python3 scripts/search-embeddings.py with each skill's description as query. High similarity between two skills (> 0.85) suggests consolidation.

2. Staleness check — For each skill:

   • When was it last modified? (git log -1 --format=%ci skills/{name}/)
   • Does it reference tools/APIs that have changed?
   • Is the description still accurate?

3. Quality gate — Spot-check frontmatter:

   • Description present and meaningful?
   • Trigger phrases included?
   • References load correctly?

4. Consolidation candidates — Skills that overlap heavily should be merged (one absorbs the other as references).

5. Deletion candidates — Skills with no clear user, stale content, or zero observations should be considered for removal.

Phase 5: Apply Changes

For each proposed change from phases 1-4:

1. Make the edit to the canonical skill/agent
2. Run ./scripts/generate-index.sh
3. Run python3 scripts/generate-embeddings.py
4. Commit with descriptive message
5. Update .spellbook.yaml if the change affects this repo's manifest

Anti-Patterns

• Adding skills because they exist, not because they fill a gap
• Keeping skills "just in case" when no one uses them
• Reviewing skills without checking consuming project observations
• Brainstorming without researching what the ecosystem already offers
• Auditing quality without checking actual usage patterns