Bootstrap

/groundrules:bootstrap

You will bootstrap a Claude Code project in the current working directory. Follow these phases in order, without skipping any. All generated files are in English (the plugin is English-only).

Phase 0 — Plugin update check (best-effort, never blocking)

1. Read version from ${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json → INSTALLED.
2. Run via Bash with a short timeout (~3s): git ls-remote --tags --refs --sort=-v:refname https://github.com/lozit/groundrules.git 'v*' | head -1
3. Extract the tag (refs/tags/vX.Y.Z). If semver-greater than INSTALLED, show an informational note (don't ask, don't stop):

   📦 groundrules vX.Y.Z is available (installed: vINSTALLED). Updating is two steps — /plugin marketplace update claude-code-groundrules refreshes the catalog but does not update the installed plugin; then reinstall it (/plugin install groundrules@claude-code-groundrules) and restart Claude Code (a new skill needs a full restart, not just /reload-plugins).

4. Fail silent: on timeout, no network, or any error, continue without mentioning the check. This is the only network access in this skill and it is best-effort (cf. ADR 0015).

Phase 1 — Scan the folder

1. Run ls -la on the cwd.
2. Detect the following markers and note what is present:
   • .git/ (existing git repo)
   • .groundrules.json (state from a previous invocation — if present, load it and switch to resume mode); a legacy pre-1.0 .starter-kit.json also triggers resume mode — load it and recommend /groundrules:migrate (V1.0 renamed the plugin and the state file)
   • CLAUDE.md, README.md, docs/, intake/, docs/media/, PLAN.md, CHANGELOG.md, .gitignore, docs/VISION.md, intake/INTENT.md
   • optional specialized docs: docs/DATA_MODEL.md, docs/SECURITY.md, docs/DESIGN_SYSTEM.md, docs/ROADMAP.md, docs/I18N.md, docs/PROCESS.md, RELEASE.md, docs/AGENT-EVALS.md
   • PLAN.md equivalents (planning aliases, same altitude) — detection is case-insensitive and nested (up to ~3 levels, excluding node_modules/.git): plan.md, TODO.md, todo.md, todos.md, TASKS.md, BACKLOG.md, including under a path (e.g. docs/gtd/todos.md). There may be several — report all of them. Case guard: never generate PLAN.md if an equivalent name exists in a different case (collision on a case-sensitive FS).
   • docs/superpowers/plans/ (superpowers per-feature plans — different altitude, not a PLAN.md alias)
   • Global / enterprise CLAUDE.md (loaded in addition to the project CLAUDE.md — NEVER overwrite it): ~/.claude/CLAUDE.md, and the managed policy if present (/Library/Application Support/ClaudeCode/CLAUDE.md on macOS, /etc/claude-code/CLAUDE.md on Linux). If at least one exists → HAS_GLOBAL_CLAUDE=true, and read its content (it's short — our own <200-line doctrine) into GLOBAL_CLAUDE_CONTENT for the content-aware tailoring in Phase 5 (which sections it already covers).
   • AI attribution policy: read (read-only) the detected CLAUDE.md files (project + global) for a rule forbidding AI attribution — patterns (case-insensitive): no AI attribution, Co-Authored-By, Generated with Claude, do not add ... attribution. If found → NO_AI_ATTRIBUTION=true. (Read-only: never modify these files.)
   • package.json (→ Node stack), pyproject.toml/requirements.txt (→ Python), Cargo.toml (→ Rust), go.mod (→ Go)
3. For each file you might create, classify it:
   • Absent → to create
   • Present with a generated-by: groundrules signature near the top → recognized, offer "ignore / regenerate". The signature may be an HTML comment (<!-- ... -->, Markdown) or a shell comment (# generated-by: groundrules, for *.sh/.gitignore), and need not be line 1 (e.g. after a shebang or YAML frontmatter — scan the first ~10 lines).
   • Present without a signature → foreign file, default "ignore"

If the folder is completely empty → classic bootstrap mode. Otherwise → resume mode (announce it clearly to the user at the start of the interview).

Phase 2 — Base interview

Group questions by theme via AskUserQuestion (max 4 questions per call for good UX).

Call 1 — Project identity

• Project name: use the folder name as the default suggestion, ask for confirmation. (1 question)
• Short description (1 sentence — the detailed intent is captured in phase 3) (1 question)
• Project type: Code (app, lib, CLI...) / Non-code (docs, research, notes...) (1 question)

Call 2 — Stack and docs (if code)

If a code project:

• Main stack: pre-fill with what you detected (Node, Python, etc.), offer Other, None in particular (1 question)
• docs/ARCHITECTURE.md: Yes (recommended for a code project) / No (1 question)
• docs/GLOSSARY.md: Yes, domain jargon / No (1 question)
• CHANGELOG.md (Keep-a-Changelog): Yes, versioned releases planned / No (1 question)

If non-code, only ask the last two (GLOSSARY, CHANGELOG).

Call 2b — Specialized docs (optional, if code)

A single multiSelect AskUserQuestion: "Which specialized docs do you want to generate? (all optional, you can check none)"

• docs/DATA_MODEL.md — entities, relationships, access rules. Check if the project has a database.
• docs/SECURITY.md — auth, access control, personal data & GDPR. Check if sensitive data.
• docs/DESIGN_SYSTEM.md — colors, typography, components. Check if the project has a UI.
• docs/I18N.md — multilingual strategy. Check if the project is multilingual.
• docs/ROADMAP.md — long-term milestone breakdown (distinct from PLAN.md).
• docs/PROCESS.md — the working-method contract: phases, validation gates, interview style. Check if the user wants a phased, gated way of working (spec → prototype → build).
• RELEASE.md — operational release runbook: environments, commands, checklists, rollback, known fragilities. Check only if the project deploys somewhere (detected CI config, hosting, or the user says so).
• docs/AGENT-EVALS.md — a log of the agent's own observed failure modes on this project (recurring mistakes, hallucinations, drifts) and the guard added for each. Distinct from LEARNINGS.md (which is about the project/domain). Offer it (unchecked by default); useful on long-running agent-driven projects.

Adapt suggestions to context: if the stack/intent suggests a UI, pre-suggest DESIGN_SYSTEM; a DB → DATA_MODEL; etc. Impose nothing: no check = no file.

Only ask this call for a code project. For a non-code project, only ROADMAP may make sense — offer it on its own if relevant.

Call 2c — Loop scaffolding (opt-in, off by default)

A dedicated single AskUserQuestion (one question), because the scaffolding is heavier than a doc and deserves its own framing: "Generate optional loop scaffolding — a maker/verifier autonomous loop (loop/) that executes loop-safe tasks with built-in back pressure? You can run it later or never; it changes nothing until you do." Options: No (recommended — add later if needed) / Yes, generate loop/. Default No.

• Skip this call entirely on resume when .groundrules.json already has loop.scaffolded: true — the namespace exists; don't re-ask or regenerate (idempotence).
• Skip this call entirely when superpowers is detected (docs/superpowers/plans/ present): superpowers already is a maker/verifier realization pipeline — groundrules defers the whole realization to it and contributes the memory layer instead. Note this in the Phase 4 recap (do not generate loop/).
• Code projects only — for a non-code project, don't ask.
• If Yes → HAS_LOOP=true. This generates the loop/ namespace (see file mapping) and inserts a ## Invariants section into the generated CLAUDE.md. If No → HAS_LOOP=false, nothing loop-related is generated.
• Brief honesty in the option/recap: the loop is most useful once /groundrules:realize lands (it will fill loop/backlog.md); for now the backlog is hand-filled. See ADR 0027 / ADR 0030.

Call 3 — Steering and git

• PLAN.md at the root (active todo maintained by Claude): Yes (recommended) / No (1 question)
• Git remote: GitHub / GitLab / None for now (1 question)
• If remote ≠ "None":
  • Visibility: Private (recommended) / Public (1 question)

Reconciliation if a PLAN.md equivalent already exists

If Phase 1 detected a planning alias (TODO.md, TASKS.md, BACKLOG.md…), do not ask the standard yes/no PLAN.md question. Ask a reconciliation question instead:

• Adopt the existing one → do not create PLAN.md; HAS_PLAN=false. Record skippedFiles["PLAN.md"] = "existing equivalent: <name>" in .groundrules.json.
• Create PLAN.md alongside → HAS_PLAN=true, create normally (the user accepts both files).
• Port the content into PLAN.md → Read the alias, create PLAN.md from the template carrying over the existing tasks (the "In progress" section), then suggest the user delete the alias. Never delete the alias automatically. Case guard: if the alias is a plan.md (same name, different case), do not create a second file — offer to adopt plan.md as-is instead.

If several equivalents are detected, list them all and clarify their respective roles (e.g. plan.md = active view, docs/gtd/todos.md = backlog) in the question. For a more complex project (brownfield, many existing docs), point to /groundrules:adopt.

superpowers: if only docs/superpowers/plans/ is present (no root alias), it is not a PLAN.md equivalent (different altitude — see the interop note in the CLAUDE.md template). Keep the standard PLAN.md question, and note in the recap (Phase 4) that PLAN.md should point to the active superpowers plan rather than duplicate tasks.

Phase 3 — Project intent

Before generating project files, capture the intent (goal, users, constraints, non-goals, acceptance criteria). It is the basis for the /groundrules:apply-best-practices skill which needs it.

AskUserQuestion: "Do you already have a brief or a vision document for this project?"

• Yes, I'll paste it now
• Yes, it's in a file (path)
• No, ask me the questions
• Skip for now

If "I'll paste it now"

Ask the user to paste the content (can be long: multiple paragraphs accepted).

Save the raw brief:

1. Read ${CLAUDE_PLUGIN_ROOT}/skills/bootstrap/templates/intake-INTENT.md.tpl
2. Substitute {{PROJECT_NAME}} and {{CONTENT}} (the brief as-is)
3. Write to intake/INTENT.md

Synthesize into a structured vision. Read the brief and extract:

• Goal: 1-3 sentences on what success looks like
• Users: who uses this
• Constraints: deadlines, budget, tech constraints, regulation
• Non-goals: what will NOT be done in V1
• Acceptance criteria: how we'll know it's done

If some points are ambiguous in the brief, ask a single grouped question to clarify the ambiguities.

If "it's in a file"

Ask for the file path (absolute or relative to the cwd). Read that file then treat it like "I'll paste it": copy into intake/INTENT.md (via the template), synthesize into docs/VISION.md.

If "No, ask me the questions"

Ask an AskUserQuestion with 4-5 questions:

• Goal: "What's the goal of the project? What does success look like?" (1-3 sentences)
• Users / personas: "Who will use this? (can be 'myself' if solo)"
• Key constraints: "What constraints (deadlines, budget, tech, regulation)?"
• Non-goals: "What will you not do in V1? (useful for scoping)"
• V1 acceptance criteria: "How will you know it's done?"

If a question is too broad, the user may answer "skip" or "to be defined later" — don't insist.

In this case, no intake/INTENT.md (the answers are already structured). Source = interview.

Generating docs/VISION.md

In all cases (except a total Skip), read ${CLAUDE_PLUGIN_ROOT}/skills/bootstrap/templates/docs-VISION.md.tpl. Substitute:

• {{PROJECT_NAME}}
• {{INTENT_SOURCE}} — intake/INTENT.md (paste) / intake/INTENT.md (file <path>) / interview
• {{GOAL}}, {{USERS}}, {{CONSTRAINTS}}, {{NONGOALS}}, {{ACCEPTANCE}} — synthesized text

Write to docs/VISION.md.

If "Skip"

Create neither intake/INTENT.md nor docs/VISION.md. Note in .groundrules.json that intent.source = "skipped". /groundrules:apply-best-practices will refuse until the vision is created.

Phase 4 — Recap and confirmation

Show a clear text recap:

• List of files that will be created (full path)
• List of files that will be ignored (resume mode)
• Planned git actions (git init, first commit, remote)

Then a final AskUserQuestion: Confirm and generate / Cancel.

Phase 5 — File generation

For each file to create:

1. Read the matching template in ${CLAUDE_PLUGIN_ROOT}/skills/bootstrap/templates/.
2. Replace the {{KEY}} placeholders with the collected values. Available placeholders:
   • {{PROJECT_NAME}} — project name
   • {{DESCRIPTION}} — short description (1 sentence, distinct from the vision)
   • {{STACK}} — stack or empty string
   • {{DATE}} — today's date in ISO (YYYY-MM-DD)
   • {{HAS_PLAN}}, {{HAS_ARCHITECTURE}}, {{HAS_GLOSSARY}}, {{HAS_CHANGELOG}} — true/false
   • {{HAS_DATA_MODEL}}, {{HAS_SECURITY}}, {{HAS_DESIGN_SYSTEM}}, {{HAS_ROADMAP}}, {{HAS_I18N}}, {{HAS_PROCESS}}, {{HAS_RELEASE}}, {{HAS_AGENT_EVALS}} — true/false (specialized docs)
   • {{HAS_LOOP}} — true/false (loop scaffolding opted in, Call 2c)
   • {{GLOBAL_CLAUDE_NOTE}} — deference note to the global CLAUDE.md + the list of omitted sections (see "CLAUDE.md generation"), or empty string if no global detected
   • {{REMOTE_PROVIDER}} — github / gitlab / empty string
   • {{REMOTE_VISIBILITY}} — private / public / empty string
   • {{CONTENT}} (intent brief) — raw brief content
   • {{INTENT_SOURCE}} — intake/INTENT.md (paste) / intake/INTENT.md (file ...) / interview
   • {{GOAL}}, {{USERS}}, {{CONSTRAINTS}}, {{NONGOALS}}, {{ACCEPTANCE}} — synthesized text
3. Plain text substitution: no template engine, just replace on each placeholder.
4. Write the file to its destination with the Write tool.

CLAUDE.md generation (content-aware tailoring against the global)

There is one template (CLAUDE.md.tpl). When a global CLAUDE.md exists, a "lean" result emerges from omitting only the sections the global actually covers — not from a separate template (cf. ADR superseding 0009).

• Priority case — CLAUDE.md project file already present (resume mode, foreign file without a signature, often tool-managed by an enterprise manager like claude-manager): generate nothing, never overwrite. Detect a free zone (marker END MANAGED / Project-Specific Notes / "below this line is yours") and offer (opt-in) to append a pointer to the groundrules docs (docs/VISION.md, docs/decisions/, docs/LEARNINGS.md, PLAN.md) — free zone only, never the managed sections. No free zone → skip. (Detailed logic: see /groundrules:adopt § "CLAUDE.md project file already present".)
• No global (HAS_GLOBAL_CLAUDE=false, no existing project CLAUDE.md) → use CLAUDE.md.tpl as-is, all sections, and substitute {{GLOBAL_CLAUDE_NOTE}} with an empty string.
• Global detected (HAS_GLOBAL_CLAUDE=true) → tailor CLAUDE.md.tpl against GLOBAL_CLAUDE_CONTENT (the global's actual content you read during detection — judge coverage, not mere presence):
  1. Collapsible sections — omit a section only if the global covers that topic's primary directive (not necessarily every line; partial coverage → bias to keep): ### Commits · ### Permissions and settings · ## Verifying the work · ## Claude Code workflow · ## Git workflow. ## Don't is always kept (it carries the signature "the repo is the only memory" line).
  2. Never omit groundrules' signature conventions even if the global mentions them — Session-start order · Capture-at-checkpoints · When-to-document routing · the-repo-is-the-only-memory · living docs · Posture; nor the project-specific — Description · Setup/Build/Test · Key files · Code/{{STACK}} · Updating-this-file · Tech stack · Notes. Bias to keep on any doubt.
  3. Drop each covered collapsible section entirely (no per-section pointer). If dropping a child ### empties its parent ## heading, drop the parent too; if a kept child remains (e.g. ### Code under ## Conventions), keep the parent.
  4. Substitute {{GLOBAL_CLAUDE_NOTE}} with the deference note + the omission list (label each omitted topic by its section title, no #):

     > **Relationship with the global CLAUDE.md**: this file is loaded **in addition to** the global (\~/.claude/CLAUDE.md` + enterprise policy) — on conflict the global/enterprise rule wins. Omitted here (your global already covers them): <comma-separated section titles, or "none">.`

  5. Recap to the user: list what you omitted and why; they can veto (keep any section).
• Net effect: a thin global → output ≈ the full template (minus only what it truly covers); a rich global → output approaches the old lean. The result scales with the global's real content, with no holes.

## Invariants section (only if HAS_LOOP=true): if loop scaffolding was opted in, read ${CLAUDE_PLUGIN_ROOT}/skills/bootstrap/templates/loop/CLAUDE-invariants.md, strip its leading <!-- ... --> comment lines, and splice its ## Invariants section into the CLAUDE.md content before the single Write — positioned immediately before the ## Posture heading (i.e. after the last child of ## Conventions, never between the ## Conventions heading and its ### subsections). Compose-then-write (bootstrap has no Edit tool); don't write-then-rewrite. If HAS_LOOP=false, do not add it (non-loop projects stay lean — the verifier is what gives invariants teeth, cf. ADR 0030). This section is never omitted by the global-tailoring logic above.

File mapping (always created)

Template	Destination
README.md.tpl	README.md
CLAUDE.md.tpl	CLAUDE.md
gitignore.minimal	.gitignore
decisions-README.md.tpl	docs/decisions/README.md
adr-template.md	docs/decisions/0000-template.md
LEARNINGS.md.tpl	docs/LEARNINGS.md
intake-README.md.tpl	intake/README.md
media-README.md.tpl	docs/media/README.md

File mapping (conditional)

Condition	Template	Destination
HAS_PLAN=true	PLAN.md.tpl	PLAN.md
HAS_ARCHITECTURE=true	ARCHITECTURE.md.tpl	docs/ARCHITECTURE.md
HAS_GLOSSARY=true	GLOSSARY.md.tpl	docs/GLOSSARY.md
HAS_CHANGELOG=true	CHANGELOG.md.tpl	CHANGELOG.md
HAS_DATA_MODEL=true	DATA_MODEL.md.tpl	docs/DATA_MODEL.md
HAS_SECURITY=true	SECURITY.md.tpl	docs/SECURITY.md
HAS_DESIGN_SYSTEM=true	DESIGN_SYSTEM.md.tpl	docs/DESIGN_SYSTEM.md
HAS_ROADMAP=true	ROADMAP.md.tpl	docs/ROADMAP.md
HAS_I18N=true	I18N.md.tpl	docs/I18N.md
HAS_PROCESS=true	PROCESS.md.tpl	docs/PROCESS.md
HAS_RELEASE=true	RELEASE.md.tpl	RELEASE.md
HAS_AGENT_EVALS=true	AGENT-EVALS.md.tpl	docs/AGENT-EVALS.md
intent.source ∈ paste/file	intake-INTENT.md.tpl	intake/INTENT.md
intent.source ≠ skipped	docs-VISION.md.tpl	docs/VISION.md
HAS_LOOP=true	loop/README.md.tpl	loop/README.md
HAS_LOOP=true	loop/maker.md	loop/maker.md
HAS_LOOP=true	loop/verifier.md	loop/verifier.md
HAS_LOOP=true	loop/LOOP.md	loop/LOOP.md
HAS_LOOP=true	loop/run-loop.sh	loop/run-loop.sh (mark executable: chmod +x)
HAS_LOOP=true	loop/backlog.md	loop/backlog.md
HAS_LOOP=true	loop/gitignore	loop/.gitignore

loop/maker.md, loop/verifier.md, loop/LOOP.md, loop/run-loop.sh, loop/backlog.md, loop/gitignore carry no {{KEY}} — copy verbatim (loop/gitignore is written to the destination loop/.gitignore). Only loop/README.md.tpl is substituted ({{PROJECT_NAME}}). loop/CLAUDE-invariants.md is not a generated file — it is the snippet inserted into CLAUDE.md (see below). Do not create loop/blocked.md or loop/lessons.md — the loop writes those on demand.

Persisted state

Write .groundrules.json at the root with this schema:

{
  "groundrulesVersion": "<current plugin version>",
  "bootstrappedAt": "YYYY-MM-DD",
  "bootstrappedWithVersion": "<current plugin version>",
  "migrations": [],
  "answers": { ... all interview answers ... },
  "intent": {
    "source": "paste" | "file" | "interview" | "skipped",
    "filePath": "<path if file>",
    "goal": "...",
    "users": "...",
    "constraints": [...],
    "nonGoals": [...],
    "acceptanceCriteria": [...]
  },
  "appliedPractices": [],
  "loop": { "scaffolded": true | false, "at": "YYYY-MM-DD" | null },
  "policies": { "noAiAttribution": true | false },
  "generatedFiles": [ ... relative paths ... ],
  "skippedFiles": { "<path>": "<reason>" }
}

policies.noAiAttribution reflects the NO_AI_ATTRIBUTION detected in phase 1 — it lets other skills (migrate, etc.) know the policy without re-scanning.

loop.scaffolded records whether the loop/ namespace was generated (Call 2c) — it lets /groundrules:realize (later) know the scaffolding exists, and keeps the opt-in idempotent (resume mode skips an already-scaffolded loop/). false/null when not opted in. The Call 2c answer lives only in this top-level loop object, not under answers (avoid a duplicate answers.loop key).

groundrulesVersion can evolve (via /groundrules:migrate), bootstrappedWithVersion stays frozen. migrations accumulates {from, to, at} entries. appliedPractices is filled by /groundrules:apply-best-practices. If the intent is skipped, intent.source = "skipped" and the other fields are null.

Phase 6 — Git

1. If .git/ absent in cwd → git init -b main.
2. git add -A
3. Check there's something to commit: git diff --cached --quiet → if nothing, skip the commit.
4. Otherwise: git commit -m "chore: bootstrap project structure with groundrules" (no version string — it would drift; the version lives in .groundrules.json and the file signatures)

AI attribution: the commit message must never contain an AI attribution marker (Co-Authored-By trailer, "Generated with Claude Code" mention, etc.). This is the bootstrap default, and it is mandatory if NO_AI_ATTRIBUTION=true — this rule overrides any default attribution guidance of the agent.

Phase 7 — Remote (if requested)

If {{REMOTE_PROVIDER}}=github:

1. command -v gh → if absent, show a manual-instruction message and go to phase 8.
2. gh auth status → if not authenticated, indicate gh auth login and go to phase 8.
3. gh repo create {{PROJECT_NAME}} --{{REMOTE_VISIBILITY}} --source=. --remote=origin --push

If {{REMOTE_PROVIDER}}=gitlab: same with glab instead of gh. The command is glab repo create {{PROJECT_NAME}} --{{REMOTE_VISIBILITY}} then git push -u origin main.

On graceful failure (CLI absent / auth missing): show the ready-to-paste command for manual execution.

Phase 8 — Final recap

Show the user:

• ✅ List of created files (relative paths)
• ✅ Git actions performed (init, short commit hash, remote URL if created)
• 📋 Suggested next steps:
  1. If the intent was captured: run /groundrules:apply-best-practices to fetch shanraisshan and apply the relevant practices to your vision.
  2. Fill intake/ with other upstream notes if relevant
  3. Flesh out CLAUDE.md (Setup/Build/Test, project-specific sections)
  4. If PLAN.md was created: add the first tasks
  5. Read docs/VISION.md and amend it if the synthesis missed something
  6. If the intent was skipped and you still want tailored practices: fill docs/VISION.md by hand then run apply-best-practices

Team-sharing tip (project-scope install)

The generated CLAUDE.md references /groundrules:* commands — but the plugin doesn't travel with git clone: a collaborator who only clones the repo won't have those commands (the model sees the instruction, the command fails). If this repo will be shared / cloned by others, suggest installing groundrules at Project scope so it's committed to .claude/settings.json and collaborators are prompted to install it on clone:

Run /plugin install groundrules@claude-code-groundrules and choose Project scope (not User). This commits the plugin reference to .claude/settings.json — anyone who clones and trusts the repo is then prompted to install it, so the /groundrules:* commands work for the whole team. (Cf. ADR 0023.)

Only surface this if .claude/settings.json doesn't already carry a project-scope groundrules entry, and don't block on it — the project's docs are plain markdown and remain readable without the plugin; only the slash-command ergonomics need it.

Important rules

• NEVER overwrite a file without explicit confirmation (see phase 4).
• Always add <!-- generated-by: groundrules v1.10.0 --> at the top of each generated file (the templates already contain it).
• Idempotence: if the user re-runs the skill, resume mode detects already-up-to-date files and does nothing.
• Surface errors: if a step fails (e.g. gh repo create returns an error), don't pretend it worked. Show the error, propose an action.
• Keep .groundrules.json: it's the source of truth for resume mode and for apply-best-practices.
• For intent synthesis: be faithful to the source text, don't invent. If the source is thin, the docs/VISION.md sections may contain "To be defined" rather than fabricated content.