CLAUDE.md — cli-code-skills

Skills installation

Skills are installed as copies in ~/.claude/skills/ for Claude Code, ~/.codex/skills/ for Codex, and ~/.jcode/skills/ for Jcode (global, available in all projects). Jcode is only touched when ~/.jcode exists. Source of truth is this repo (~/workspace/cli-code-skills/).

After modifying a skill in this repo, copy it to every runtime:

scripts/install_skills.sh

Or install only one runtime:

scripts/install_skills.sh --claude
scripts/install_skills.sh --codex
scripts/install_skills.sh --jcode

The installer removes each destination skill before copying it. That avoids stale files when a reference is renamed or deleted.

The safe direction is repo → runtime. If a runtime copy was edited in place and that edit is not yet in the repo, the installer refuses to overwrite it (exit 3) rather than destroy it. It tracks this with a .skills-manifest written into each runtime skills root at install time.

scripts/install_skills.sh --check          # report divergence, change nothing
scripts/install_skills.sh --pull-claude    # copy Claude in-place edits back into the repo
scripts/install_skills.sh --pull-codex     # copy Codex in-place edits back into the repo
scripts/install_skills.sh --pull-jcode     # copy Jcode in-place edits back into the repo
scripts/install_skills.sh --force          # overwrite even runtime-ahead edits

If you (or a session) edited a skill directly in ~/.claude/skills/, run --pull-claude to bring it back into the repo, review git diff, validate, and commit — then a normal install re-syncs both runtimes.

Shared files (gotchas.md and shared/) are referenced by root SKILL.md files as ../<name> and by references/ files as ../../<name>. The installer copies them to the skills root for every runtime.

Do NOT use symlinks — use real copies.

Validation

Before committing or installing changes, run:

python3 scripts/validate_skills.py

The validator checks frontmatter, skill counts, README/CLAUDE inventories, shared-file paths, and Claude/Codex-neutral cli-cycle orchestration.

It also enforces the Agent Skills spec limits: name <= 64 chars and lowercase/digits/hyphens only, description <= 1024 chars (over the limit the skill is rejected by the API). SKILL.md bodies over 500 lines are reported as WARN, following Anthropic's progressive-disclosure guidance: move detail into references/ so the body stays cheap to load.

Evaluations

evals/ holds one JSON case per scenario plus the deliberately flawed fixtures they point at. There is no scored runner: grading a skill run needs a model in the loop. What is automated keeps the cases honest.

python3 scripts/check_evals.py

It rejects a case whose skill or fixture does not exist, or that cites a dimension ID the skill never defines. See evals/README.md. Fixtures are broken on purpose and must not be repaired.

Current skills (36)

Audit (12): cli-audit-code, cli-audit-data, cli-audit-doc, cli-audit-drift, cli-audit-hanoi, cli-audit-review, cli-audit-shell, cli-audit-sync, cli-audit-tangle, cli-audit-test, cli-audit-wizard, cli-audit-xray

Cycle (1): cli-cycle

Forge (21): cli-forge-arch, cli-forge-chef, cli-forge-choice-ux, cli-forge-data, cli-forge-demo, cli-forge-doc, cli-forge-github, cli-forge-hld, cli-forge-infra, cli-forge-lld, cli-forge-oci-rootless, cli-forge-perf, cli-forge-pipeline, cli-forge-plume, cli-forge-prez, cli-forge-profile, cli-forge-quorum, cli-forge-readme, cli-forge-resilience, cli-forge-schema, cli-forge-tree

Git (1): cli-git-conventional

Watermark (1): cli-watermark

Git identity

Use clement <cliard@a00.fr> for this repo. Never use naval-group email.

Commit style

This repo uses the cli-git-conventional skill. Never add Co-Authored-By trailers. Follow Conventional Commits v1.0.0 spec. Ghostwriter style — human voice, no AI markers.