generate-do-work

Generate Do-Work

Scaffolds a customised do-work skill into the current repo from templates/do-work.template.md (relative to this skill).

Workflow

1. Auto-detect repo specifics

Run detection, do NOT ask user for anything inferable. Gather:

Field	How to detect
REPO_NAME	basename of git toplevel (git rev-parse --show-toplevel)
pkg manager	lockfile: pnpm-lock.yaml→pnpm, yarn.lock→yarn, package-lock.json→npm, bun.lockb→bun
monorepo packages	pnpm-workspace.yaml / package.json#workspaces → list package dirs + names
lint / type-check / test cmds	package.json#scripts — grep keys for lint, lint:ts/typecheck, test
language rules	tsconfig.json (strict, ESM via type:module), presence of .ts
data sink	scan for queue/db usage (prisma, sqs, kafka) → set {{SINK}}; default "DB/queue"
commit hook	.husky/commit-msg, commitlint.config.*, .commitlintrc*

Use one git/file batch, not many Read calls. Prefer reading package.json, lockfile names, workspace config, and any husky/commitlint files.

2. Resolve commit-message section (important)

Two cases:

• Husky + commitlint present (e.g. .husky/commit-msg runs commitlint and auto-prepends a Jira ticket from the branch name):
  • Emit Conventional Commits body only.
  • Set COMMIT_HOOK_NOTE = — the commit-msg hook prepends the Jira ticket from the branch name automatically, so do NOT add it manually.
  • Set COMMIT_RULES_EXTRA = - Ticket key is added by the husky hook from the branch name; if branch has no ticket, prepend \JIRA-1234: ` manually.`
• No hook: emit Conventional Commits, and add a rule that the user should prepend the Jira key manually (JIRA-1234: <type>(<scope>): ...) if their team uses Jira-first history. Detect Jira-first by sampling git log --oneline -30 for a ^[A-Z]+-\d+ prefix.

3. Fill placeholders

Replace every {{...}} token in the template:

• {{REPO_NAME}} — repo name
• {{PACKAGES_NOUN}} — "packages" (monorepo) or "modules/files" (single pkg)
• {{PACKAGES_LIST}} — (pkg-a, pkg-b, ...) or empty
• {{PACKAGE_SINGULAR}} — "package" or "area"
• {{PATTERNS_REF}} — (see CLAUDE.md) if a CLAUDE.md exists, else empty
• {{LANG_RULES}} — bullet lines for the detected language (TS strict/ESM/workspace:*, etc.); empty if none
• {{SINK}} — detected data sink or "DB/queue"
• {{FEEDBACK_COMMANDS}} — actual shell commands in order (type-check, lint, test) using the detected pkg manager + script names
• {{FEEDBACK_NOTES}} — autofix hint if a lint:fix script exists, else empty
• {{COMMIT_HOOK_NOTE}}, {{COMMIT_RULES_EXTRA}} — from step 2
• {{SCOPE_HINT}} — "package name or domain" (monorepo) or "feature/domain area"

Leave no {{...}} unresolved. Empty placeholders → remove the line if it becomes blank/dangling.

4. Confirm gaps only

If any field could NOT be inferred (e.g. no test script, ambiguous lint cmd), ask the user only those in one batched question. Never re-ask what was detected.

5. Write the skill

Write resolved content to .claude/skills/do-work/SKILL.md at repo root (create dirs). Use the Write tool — never shell redirection.

Then report: detected values table + output path. Tell user to /do-work (or restart session if skills are session-loaded).

Notes

• Front-end repos: keep the front-end block, drop or de-emphasise TDD loop.
• Idempotent: regenerating overwrites the existing do-work skill.