📚 A curated collection of skills for AI agents — turning tacit knowledge into reliable execution.

English | 中文

npx skills add lencx/skills

# Claude Code
npx skills add lencx/skills -a claude-code -g

# Codex
npx skills add lencx/skills -a codex -g

• keel: Architecture design and governance protocol for system structure, boundaries, contracts, migrations, and long-lived codebase health. Keeps the load-bearing spine small, owned, checked, and deletable over time.
• coding-protocol: Ambient, risk-scaled execution protocol for coding tasks (ref: Andrej Karpathy's public observations).

Skills follow the open Agent Skills format (SKILL.md with name/description frontmatter) and reference no agent-specific tools, so they work with any agent that reads the format — Claude Code, Codex, Cursor, and others. npx skills add handles per-agent installation.

Tip: these skills pair best with declarative prompts — state the success criteria ("make these tests pass") rather than step-by-step instructions.

• node scripts/validate-skills.mjs — two-layer validation: universal checks for every skill (frontmatter, size limits, referenced files, section cross-references), plus inferred archetype packs (protocol: rules ↔ failure modes ↔ evals coverage). Runs in CI on every push and PR.

• evals/ — project-level behavioral A/B scenarios. evals/<skill>.md is matched to skills/<skill>; evals/manifest.json stores only overrides and project-level exceptions such as cross-skill collaboration boundaries. Use archetype: "none" only for skills intentionally outside an inferred pack.

• Bump the skill's SKILL.md version whenever installed skill content changes (SKILL.md or files under that skill directory).

• Local development with Claude Code: symlink the skill so the installed copy always tracks the repo (re-run npx skills add instead if you prefer copies):

  ln -sfn "$(pwd)/skills/keel" ~/.claude/skills/keel
  ln -sfn "$(pwd)/skills/coding-protocol" ~/.claude/skills/coding-protocol

MIT