Skip to content

CLAUDE.md

Latest commit

 

History

History
35 lines (26 loc) · 1.83 KB

CLAUDE.md

File metadata and controls

35 lines (26 loc) · 1.83 KB

CLAUDE.md — skill-forge

Read CONTEXT.md for project vocabulary. Read AGENTS.md for available skills and scripts.

Identity

This is a self-improving skill collection with two objectives: (1) make itself better each cycle, (2) produce skills that earn stars. The meta-agent in root SKILL.md orchestrates discovery, creation, routing, and improvement.

Key Facts

  • Skills are in skills/ directory, each with a SKILL.md
  • Memory/state persists in memory/*.json (indexed via scripts/index-memory.js)
  • Quality validation: node scripts/validate-skill.js skills/<name> — must pass with 0 warnings
  • Self-check runs after every action (Phase 5 in root SKILL.md)
  • Self-improvement: node scripts/skillopt.js tracks outcomes and proposes prompt deltas
  • Routing: node scripts/route-task.js "task" — TF-IDF semantic matching with compound skills
  • Publishing target: GitHub under Adit-Jain-srm org
  • Install: npx skills@latest add Adit-Jain-srm/skill-forge

When Modifying

  • Never ship a skill without validation passing (0 warnings)
  • Run node tests/run-tests.js after any structural change (must pass all 98 tests)
  • Always update CONTEXT.md if adding new domain terms
  • Learnings go in memory/learnings.json and must change BEHAVIOR not just exist
  • After using a skill, record outcome: node scripts/skillopt.js --record --skill <name> --task "..." --outcome <good|poor|mixed>
  • README frames skills as PROBLEMS not features

Reasoning Approach

When making decisions about skill creation, routing, or improvement:

  1. State the specific problem being solved (not vague goals)
  2. Consider what already exists (check memory/learnings.json and indexed-learnings.json)
  3. Validate against the best (not "good enough" — compare to top performers)
  4. Prove the result works (run tests, show evidence, never claim without verification)