Commit .claude skills to the repo so the agent setup is portable

[vivek7405]

Problem

The repo's .claude/ ships the HOOKS (.claude/settings.json + block-prose-punctuation.sh / nudge-uncommitted.sh / require-tests-with-src.sh / route-skills.sh), so a fresh git clone on another computer gets the guardrails. But the SKILLS those hooks depend on (webjs-file-issue, webjs-start-work, webjs-list-todos, use-railway) live ONLY in the maintainer's machine-local ~/.claude/skills/ and are NOT committed. So on another computer the setup is half-broken:

• The skills do not exist, so /webjs-start-work, /webjs-file-issue, etc. are unavailable.
• Worse, the committed route-skills.sh keyword-matches prompts against those exact skill names and INJECTS a directive to "invoke before other work". On a clone where the skill is absent, the hook points the agent at a skill that does not exist.

The hooks and the skills they route to are split across two homes (repo vs ~/.claude), so the repo is not self-sufficient.

Design / approach

Commit the skills into the repo's .claude/skills/<name>/SKILL.md (Claude Code loads project-level skills from there), so a clone is self-sufficient and route-skills.sh's references resolve. This fits the repo's existing AI-first posture (it already commits hooks + AGENTS.md + the scaffold's per-agent configs).

• The three webjs-* skills are webjs-project-specific (the project board, the start-work lifecycle), so they belong in the repo.
• use-railway is general-purpose; commit a copy too so the repo's deploy-verification workflow is portable, but it can also remain a personal skill for other repos.
• Drift: a skill living in both ~/.claude/skills/ and the repo can diverge. The repo copy is canonical for webjs work. SAFE DEFAULT for this issue: commit COPIES (additive, non-destructive); do NOT delete the maintainer's ~/.claude copies as part of this change. A follow-up can decide whether to remove the home copies of the webjs-* ones or symlink them, since that affects the maintainer's other repos.

Implementation notes (for the implementing agent)

• Where to put them: .claude/skills/webjs-file-issue/SKILL.md, .claude/skills/webjs-start-work/SKILL.md, .claude/skills/webjs-list-todos/SKILL.md, .claude/skills/use-railway/ (the last has a references/ + scripts/ tree; copy it whole). Source is ~/.claude/skills/<name>/.
• The committed router .claude/hooks/route-skills.sh already references all four skills by name; once they are committed the references resolve on a clone. Re-read it to confirm no path assumptions break when the skills are project-level rather than user-level.
• Skill precedence: project-level and user-level skills can both be present on the maintainer's machine. Verify Claude Code does not error or double-load when a skill name exists in BOTH ~/.claude/skills/ and .claude/skills/ (it should prefer/merge, but confirm there is no conflict; this is why the safe default is copies + a documented canonical).
• use-railway is a multi-file skill (SKILL.md + references/*.md + scripts/*.py + scripts/railway-api.sh). Copy the entire directory, not just SKILL.md.
• Tests / validation surface: there is a test/hooks/route-skills.test.mjs (per AGENTS.md). Confirm it still passes, and extend it if it should assert the routed skills now exist in-repo.
• Docs: note in the repo AGENTS.md (the "Skills are routed deterministically" section) that the skills are now committed under .claude/skills/, so the routing is self-contained.

Acceptance criteria

• .claude/skills/{webjs-file-issue,webjs-start-work,webjs-list-todos,use-railway} are committed (use-railway with its full references/ + scripts/ tree).
• A fresh clone has working /webjs-* + /use-railway skills, and route-skills.sh's name references all resolve (no dangling skill).
• test/hooks/route-skills.test.mjs passes; extended if appropriate to assert the in-repo skills exist.
• No secrets are committed (the skills reference the public project board + railway telemetry caller, no tokens).
• The maintainer's ~/.claude copies are left intact (non-destructive); drift handling is documented as a follow-up.
• AGENTS.md updated to note the skills are now in-repo.