dotbabel

# CLAUDE.md — Global Claude Code Rules

<!-- dotbabel:cli claude -->

> **Two ways to use dotbabel:**
>
> - **TL;DR — just want skills & commands:** Clone this repo and run `./bootstrap.sh`.
>   That wires `commands/`, `skills/`, and this file into `~/.claude/` in one step.
>   No npm required.
> - **Want more:** Install `@dotbabel/dotbabel` for the full CLI — `dotbabel bootstrap`,
>   `dotbabel sync`, `dotbabel doctor`, `dotbabel init`, `dotbabel detect-drift`,
>   and the spec-governance validators.
>   See [README.md](./README.md) or [docs/quickstart.md](./docs/quickstart.md).
>
> **This file is for the bootstrap path.** It gets symlinked into `~/.claude/CLAUDE.md`
> by `bootstrap.sh` and sets the global rule floor for every Claude Code session.
> **Consumers of `@dotbabel/dotbabel` do NOT inherit it** — the plugin's behavior
> lives in `plugins/dotbabel/` and [docs/](./docs/). Contributors should read
> [CONTRIBUTING.md](./CONTRIBUTING.md) first.

<!-- dotbabel:end -->

Universal behavior for every Claude Code session in every repo. Project-level `CLAUDE.md` files extend and may override these, but should not repeat them.

<!-- dotbabel:rule-floor:begin -->

## Local filesystem conventions

- All projects live at `$HOME/projects/`. Do not search the home directory or default locations.
- Global Claude config lives wherever you cloned `dotbabel` and is symlinked into `~/.claude/`. Edit files in the clone, not `~/.claude/` directly.

## Disk hygiene

- The Windows C: drive on this machine is small. The WSL virtual disk only grows; it never shrinks on its own. Every gigabyte you write inside the distro stays allocated on C: until a manual compact.
- After you run containerized tests (testcontainers), run `docker volume prune -f`. Anonymous test volumes leak ~70 MB per run and are the main growth driver.
- Keep large artifacts, datasets, and clones on `/mnt/wsl/storage`, not on the distro disk.
- Do not download large toolchains or browser bundles (Playwright browsers, Go module trees) unless the task needs them. Prefer existing installed versions.
- Cleanup is automated on the Windows side: scheduled tasks `WSL Weekly Maintenance` and `WSL Monthly Compact` run `C:\WSL\wsl-maintenance.ps1` (log: `C:\WSL\maintenance.log`). Do not build ad-hoc cleanup scripts; extend that one.

## Code Changes

- Before proposing fixes, **read the relevant source files**. Use `Grep` + `Glob` + `Read` to locate current behavior.
- Cite `file:line` references in every analysis. Claims without citations are not grounded.
- Do not propose edits until the analysis is confirmed against real code. "The file is probably named X" is not grounding — open it.
- When unsure, invoke the `/ground-first` skill to enforce the read-first discipline.
- **Surface assumptions before coding.** If a request has multiple valid interpretations, list them explicitly. In interactive sessions, ask before picking one. In autonomous/headless mode, state the chosen interpretation and proceed. "Make it faster" → clarify which dimension (latency, throughput, perceived UX) before writing code.
- **Surgical orphan cleanup.** When your changes make an import or variable unused, remove it. Remove a function only after verifying it is not part of a public/exported API and has no remaining references (use a repo-wide search); otherwise keep it or deprecate it. Don't remove pre-existing dead code your changes didn't create — mention it instead.

## Root Cause Before Fix

- For any bug or data discrepancy, perform a grounded audit (read the actual code paths, check deployment state, verify data sources) BEFORE proposing a fix or plan. Do not accept the first plausible hypothesis.
- State evidence (file:line, log snippet, commit sha) for each claim in the diagnosis.
- Present at least two candidate root causes with evidence for and against each before settling on one.
- **Do not write code until the user approves the audit.** In interactive sessions, wait for explicit sign-off. In autonomous/headless mode, emit the audit and state the chosen root cause before proceeding.

## Testing

- Run the project's **full** test suite locally before merging any PR that modifies a file matching `critical_paths` (see the `quality` key in `.dotbabel.json`) or anything consumed by downstream consumers.
- Never claim a test failure is "pre-existing" without proving it. Required proof:
  ```bash
  git stash && <test-command> ; git stash pop
  ```
  If the failure survives the stash, it's pre-existing. If it disappears, your change introduced it.
- Detect the test runner from the project, don't guess:
  - `Makefile` with a `test` target → `make test`
  - `package.json` → `npm test` (or `pnpm test` / `yarn test` based on the lockfile)
  - `go.mod` → `go test ./...`
  - `pyproject.toml` → `pytest` or `uv run pytest`
- Partial test subsets are fine for iteration. Full suite is required before pushing or merging.

## TDD and verification

- **Always follow TDD for new features:** write tests first (positive, negative, boundary), then implement until tests pass.
- **For bug fixes:** write a failing test that reproduces the issue, fix, then verify.
- **Transform vague tasks into verifiable goals before starting.** "Fix the bug" → "write a test that reproduces it, then make it pass." For multi-step tasks, emit a concise plan with explicit verification at each step: `Step → verify: [check]`. Default to 5 bullets or fewer; exceed that only when the task is genuinely complex.
- **When editing Go files, run `gofmt -w <file>` immediately after editing.** Never leave Go files with formatting issues.
- **When reporting status or roadmap progress, verify each item against actual code or config before marking it complete.** Do not assume completion — show the evidence.

## Test Plan Verification

- Run every command in the test plan verbatim, in order. Paste the **last 10 lines of output** for each.
- If any command was skipped or inferred rather than run, say so explicitly. Never claim completion based on partial runs.

## Version control discipline

- **Never push to `main` (or any branch) without explicit user instruction.** Commit locally and wait for the user to say "push".
- **Never merge a PR without explicit user instruction.** Do not use `--auto`, `gh pr merge`, or any merge path unless the user says "merge" for that specific PR.
- **Never force-push, force-rebase, or `git reset --hard` a branch that is not yours.** If conflict resolution is ambiguous, stop and ask.
- **Never undo or revert another session's committed work.** Prior session commits are authoritative. If a merge conflict arises with prior session work, stop and ask.
- Before pushing any commit, review staged files for sensitive content (.env, credentials, API keys). Use `.gitignore` proactively.
- Prefer new commits over `--amend`. Never pass `--no-verify` or `--no-gpg-sign` unless the user explicitly asks.

## Worktree discipline (for any non-trivial change)

- **Default to git worktrees for anything non-trivial.** New features, bug fixes, code reviews, refactors, and spec work belong in a fresh worktree under `.claude/worktrees/<slug>/`, branched from the latest `origin/main` (run `git fetch origin main` first).
- The main checkout is effectively read-only for agentic work unless the user says "do it on main" for this specific task. A one-line typo fix they want committed directly is fine; anything larger is not.
- Never use `gh pr checkout`, `git checkout <other-branch>`, `git switch`, or `git stash` in the main checkout as a way to swap contexts; those operations silently corrupt any concurrent session editing the same checkout.
- **Respect other sessions' worktrees and branches.** Multiple agents and humans work concurrently. Before creating a worktree, run `git worktree list` and scan for anything that looks active (recent HEAD, branch name matching your intent). Never remove, rename, or force-overwrite a worktree you did not create in this session.
- **Clean up your own worktree when the work lands.** After the PR merges, or after you abandon the task, remove the worktree you created in this session: `git worktree remove .claude/worktrees/<slug>` from the main checkout, then `git worktree prune`. This deletes only the checkout directory — the branch and its commits survive. Do it before you end the task. No other session will do it for you, because the rule above forbids them from touching a worktree they did not create.
- **Never remove a worktree that holds content living nowhere else.** Run `git status --porcelain` first. Uncommitted edits to tracked files, and untracked files absent from `origin/main`, are destroyed by removal — commit them to the branch or ask the user before you remove. Regenerated exports, coverage output, build artifacts, and `test-results/` are safe to discard.

## Worktree & Sandbox Conventions

- Before starting work in a worktree, verify it is clean (`git status`) and not already claimed by a concurrent headless worker (check for lockfiles/PID files).
- Use `$CLAUDE_PROJECT_DIR` in hooks and scripts rather than relative paths.
- When sandbox blocks writes to `/tmp` or the worktree path, emit results to stdout as a fallback and flag the limitation explicitly.

## PR Conventions

- Create PR bodies via `gh pr create --body-file <file>`, not heredoc. Heredocs mangle backticks and break the required Spec ID block.
- Required sections in every PR body:
  - `## Summary` — 1–3 bullets describing the change.
  - `## Test plan` — bulleted markdown checklist.
  - `## Spec ID` heading followed by the spec id — if the project uses spec IDs (check for `specs/` or `docs/specs/`). Must be an H2 heading; `dotbabel-check-spec-coverage` extracts it via H2 regex.
- **The `## Spec ID` section must contain nothing but the id(s).** The extractor captures everything from that heading to the next H2 heading _or the end of the body_, then splits it on whitespace and treats every token as a spec id. Anything trailing — a generated-by footer, a sign-off, a link — is parsed as unknown spec ids and fails the gate. Put `## Spec ID` last with nothing after it, or follow it with another H2. `## No-spec rationale` is exempt — its body is only checked for non-emptiness, never tokenized — so prose and trailing content are safe there.
- Never merge a PR with failing CI without explicit user approval.

## Shell & Scripting

- Use `bash` (not `zsh`) for monitor scripts, loops, and anything using `read`, `$?`, or `$status`. `zsh` makes `status` read-only and breaks scripts silently.
- Avoid reserved variable names: `status`, `path`, `pwd`, `prompt`, `HISTFILE`. Prefer `result`, `workdir`, `current_status`.
- Before long-running work, verify session sanity:
  - `pwd` exists (sessions die silently on deleted worktrees).
  - `git status` is clean (or intentionally dirty) — no unexpected locks.
  - The branch is what you expect.
- Prefer `gh <cmd> --body-file` or `--json` + `--jq` over shell-interpolated strings.

## Deploy discipline

- **Never deploy to production without explicit user instruction.** Use the project's sanctioned deploy command (e.g. `/ship`, not direct `vercel --prod` or `flyctl deploy`).
- **When designated as autonomous** (batch task, pipeline, overnight run), do not stop for permission at intermediate steps. Execute fully. Only pause for genuinely destructive or irreversible actions.
- **Autonomous dry-run contract.** Before invoking any command that writes to production data, emit a one-block plan: exact command, every flag with a justification, expected scope, estimated runtime. Then execute without further prompts. Never pass `--force` without explicit user authorization for the specific run.

## Implementation vs Spec

- When the user asks for an implementation, a fix, a PR, or "just do X" — **cap planning at a 5-bullet sketch, then edit**. Do not spin up spec docs.
- Use `/spec` only when the user explicitly asks for a spec, design doc, RFC, or says "let's spec this out."
- If a task genuinely needs a plan longer than 5 bullets, write it inline in the response — don't create a planning file unless asked.

## Headless Mode

For recurring sweeps (Dependabot, cron, CI-triggered agents), use headless mode to skip tool-approval prompts.

<!-- dotbabel:cli claude -->

```bash
claude -p "Check rebase status of all open Dependabot PRs and report CI status" \
  --allowedTools "Bash(gh:*),Read,Grep"
```

Scope `--allowedTools` tightly — prefer `Bash(gh:*)` over `Bash(*)`. Combine with cron or GitHub Actions for unattended runs.

<!-- dotbabel:end -->

## AI code quality floor

Use the resolved project policy. `dotbabel quality` measures it with the tools the repository already has. It never installs a checker.

- Run `dotbabel quality explain` before a policy-sensitive change. Add `--rule <id>` for one rule.
- Run `dotbabel quality detect` to see the components, the selected tools, and the trust state. It executes no project command.
- Run `dotbabel quality check --profile fast` while you edit. Run `--profile pr --base <ref>` before a pull request. Keep `--profile deep` for a scheduled audit.
- Add `--path <glob>` to narrow a run to one package. Add `--all` to check the whole repository instead of a diff.
- Read the exit code. `0` is no error verdict. `1` is a policy failure. `2` is a missing tool, report, base, or trust. `64` is invalid usage. Never report `2` as a pass.
- Set project policy in the repository `.dotbabel.json` under the `quality` key. Do not lower a shipped threshold to make a check pass.
- Trust project commands by their exact repository path. Pass `--allow-project-commands` for one CI run only.
- Resolve an ambiguous tool choice in configuration. Do not guess, and do not install a missing tool.
- Simplify control flow before splitting a function. Split a file only when each result has one coherent responsibility.
- Add tests for behavior and failure boundaries. Reject assertion-free or implementation-coupled coverage padding.
- Do not add abstractions only to reduce local metrics. Remove obsolete code instead of moving it.
- Do not replace a safe `unknown` value with a cast. Narrow or validate untrusted data before use.
- Treat each new suppression as a review finding. Keep each exception narrow, temporary, and justified.
- Exclude generated and vendor code only with file evidence. Do not exclude difficult code for convenience.
- Use changed-code checks and no-regression rules in legacy repositories. Report improvements that remain above a target.
- Report unsupported, unavailable, and not-configured measurements. Never claim that an unsupported metric passed.

## Communication

**Hard caps. Not aspirational — enforced.**

- Default response is ≤3 sentences. Prose, not bullets. No headers.
- Never restate the question. Never preface with "Let me…", "I'll…", "Here's…", "Looking at…", "Based on…". Start with the answer.
- Never summarize what you just did at end-of-turn. The diff and tool output already show it. One line max if a follow-up genuinely matters; otherwise zero lines. Exception: the status report in the next rule.
- **At the end of each work unit, ALWAYS give a completion status report** of the feature, spec, or plan in progress. A work unit is a task, a commit, a PR step, or a turn that changes state. Give what is done, what remains, and the completion count or percentage, verified against actual code or config. Then give the next optimal step as one suggestion. Keep the report short (≤5 lines). This rule overrides the length caps in this section.
- No bullet lists unless the answer is genuinely ≥3 peer items. Two items = a sentence with "and".
- No headers (`##`, `###`) in chat responses. Headers belong in files, not conversation.
- Tool-use narration: one short sentence per _meaningful_ step (found the bug, changing direction, blocked). Silent for routine reads/greps.
- No hedging filler: "essentially", "basically", "it's worth noting", "to be clear", "I should mention", "keep in mind", "as you can see".
- No closing affirmations: "Let me know if…", "Happy to…", "Hope this helps", "Feel free to…". Just stop.
- Code answers: show the code. Skip the prose explanation unless asked. If the user wants the reasoning they'll ask.
- When in doubt: cut it. A terse answer the user re-asks for detail on beats a wall they have to skim.

**Length rubric.** Simple factual question → one sentence. Code change → the diff + ≤1 line of context. Investigation result → ≤3 sentences + file:line. Spec/architecture discussion → as long as needed, but earn every paragraph.

## Language: ASD-STE100 Simplified Technical English

**Write all chat output to the user in ASD-STE100 Simplified Technical English (STE).** This rule applies to every AI agent that reads this rule floor.

- Use approved STE words where the dictionary permits. Use one word for one meaning.
- Write short sentences. Use a maximum of 20 words in an instruction. Use a maximum of 25 words in a description.
- Use the active voice. Use the present tense where possible.
- Give one instruction in each sentence. Start each instruction with a verb.
- Write paragraphs with a maximum of 6 sentences.
- Do not use idioms, slang, or Latin abbreviations such as "e.g." and "i.e.".
- Use the articles "a", "an", and "the". Do not remove them.
- Keep technical names, code, file paths, commands, and quoted output as they are. STE does not change code.
- Put safety warnings before the instruction they apply to.
- The Communication hard caps stay in effect. STE controls the style; the caps control the length.

## Protected paths (dogfood)

This repository governs itself with `@dotbabel/dotbabel`. The authoritative
list of protected paths lives in `docs/repo-facts.json` and every entry must
be documented in every rule-floor file listed in
`docs/repo-facts.json:rule_floor_files`; `dotbabel-check-instruction-drift`
enforces this invariant.

- `CLAUDE.md` — canonical rule-floor source.
- `README.md` — top-level public README.
- `AGENTS.md` — project-scoped instructions for Codex / Copilot / OpenCode CLI.
- `GEMINI.md` — project-scoped instructions for Gemini / Antigravity CLI.
- `.github/workflows/**` — CI pipelines.
- `.github/copilot-instructions.md` — project-scoped instructions for GitHub Copilot.
- `.claude/**` — skill manifest, settings, hooks.
- `docs/repo-facts.json` — the facts source of truth.
- `docs/specs/**/spec.json` — spec metadata governed by the spec-anchored workflow.
- `plugins/dotbabel/src/**` — the npm package's source of truth.
- `plugins/dotbabel/bin/**` — the shipped bin entrypoints.
- `plugins/dotbabel/templates/**` — scaffolding templates consumers install (includes `plugins/dotbabel/templates/cli-instructions/**`, the user-scope rule-floor templates generated by `dotbabel-generate-instructions`).

Any PR touching one of these paths must carry either `Spec ID: dotbabel-core`
or a `## No-spec rationale` section in its body.

The rule-floor block (between `<!-- dotbabel:rule-floor:begin -->` and `<!-- dotbabel:rule-floor:end -->` markers) in `AGENTS.md`, `GEMINI.md`, and `.github/copilot-instructions.md` is **auto-generated from this file** by `dotbabel-generate-instructions`. Edit the rule floor here in `CLAUDE.md`; re-run the generator (`npx dotbabel-generate-instructions` or `dotbabel sync`) to fan it out. Hand-editing the block in a host file will be reverted by the next regen and is detected by `dotbabel-check-instruction-drift`.

## Skills, Commands, and Discovery

<!-- dotbabel:cli claude -->

Claude Code now treats skills and custom commands as the same slash-invoked
family: `commands/foo.md` and `skills/foo/SKILL.md` both expose `/foo`.
Skills are the preferred shape for new reusable workflows because they support
supporting files, richer frontmatter, automatic model invocation, direct user
invocation, and lazy-loaded references. Existing `commands/` files remain valid;
do not migrate them just for naming.

Use `skills/<id>/SKILL.md` when:

- Natural-language requests should route to the workflow, not only `/id`.
- The workflow needs `references/`, scripts, templates, examples, or other
  supporting files.
- Invocation behavior matters: `disable-model-invocation`, `user-invocable`,
  `allowed-tools`, `model`, `effort`, or `context` belongs in skill frontmatter.
- The capability should appear in the generated taxonomy as a reusable skill.

Use `commands/<name>.md` only for simple existing single-file prompt templates
where explicit slash invocation is the whole interface and no supporting files
or automatic routing are wanted. For side-effectful workflows implemented as
skills, set `disable-model-invocation: true`; do not fall back to commands
solely for safety.

`handoff` intentionally stays a skill: phrases like "continue in codex" and
"push handoff" should auto-route, while `/handoff` still works as a direct
invocation.

`pr-conductor` is side-effectful but model-invocable on purpose: an agent can
start it after it commits a branch. The shipped settings put
`Skill(pr-conductor)` in `permissions.ask`, so each agent-started run waits for
the user to approve the Claude Code prompt.

Loading model: skill descriptions are available for discovery, full `SKILL.md`
content loads only when invoked, and supporting files load only when referenced.

<!-- dotbabel:end -->

Do not maintain static command or skill tables in instruction files. When editing
this dotbabel repository, the authoritative inventory is generated from artifact
frontmatter:

```bash
node plugins/dotbabel/bin/dotbabel-index.mjs --check
node plugins/dotbabel/bin/dotbabel-list.mjs --type skill
node plugins/dotbabel/bin/dotbabel-list.mjs --type command
node plugins/dotbabel/bin/dotbabel-search.mjs <query>
node plugins/dotbabel/bin/dotbabel-show.mjs <id> --type skill
```

<!-- dotbabel:rule-floor:end -->

More agent context in kaiohenricunha/dotbabel

48 other files this repository gives its agents.