OrodruinLabs/nazgul:bootstrap-project

Bootstrap Project

Examples

• /nazgul:bootstrap-project — Interactive: ask for objective, then run full pipeline
• /nazgul:bootstrap-project "Add Stripe billing to existing app" — Use given objective
• /nazgul:bootstrap-project --dry-run — Run pipeline and transform into scratch without relocating
• /nazgul:bootstrap-project --yes --overwrite — Non-interactive, overwrite existing ./docs or ./.claude/agents

Arguments

$ARGUMENTS

Flags

• --yes — Non-interactive; abort on ambiguous prompts.
• --overwrite — Force overwrite of non-empty ./docs/ or ./.claude/agents/.
• --dry-run — Run pipeline and transform into scratch; skip relocation and cleanup.
• --wipe-scratch — Delete any existing ./.bootstrap-scratch/ before starting.
• --resume-scratch — Keep any existing ./.bootstrap-scratch/ and proceed with the run. Pipeline agents are still invoked; each may reuse pre-existing files in scratch at its own discretion, but the skill does not explicitly skip phases based on what's already there.

Instructions

Pre-load: Run ToolSearch with query select:AskUserQuestion to load the interactive prompt tool (deferred by default). Do this BEFORE any step that uses AskUserQuestion.

Phase 1 — Pre-flight gate

Source the preflight helpers:

source "$CLAUDE_PLUGIN_ROOT/scripts/lib/bootstrap-preflight.sh"

Parse $ARGUMENTS structurally: split on whitespace, pull out known flags, and leave the rest as the free-form objective. This prevents false positives where an objective string contains a flag-like substring (e.g. "document the --dry-run workflow").

BOOTSTRAP_YES=false
BOOTSTRAP_OVERWRITE=false
BOOTSTRAP_DRY_RUN=false
BOOTSTRAP_WIPE_SCRATCH=false
BOOTSTRAP_RESUME_SCRATCH=false
BOOTSTRAP_OBJECTIVE=""

# Disable pathname expansion during tokenization so an objective containing
# glob metachars (*, ?, [) doesn't expand to matching filenames. Re-enable
# afterwards so downstream code behaves normally.
set -f
# Intentionally NOT quoted: we want word-splitting so each token is processed.
# shellcheck disable=SC2086
set -- $ARGUMENTS
for tok in "$@"; do
  case "$tok" in
    --yes)           BOOTSTRAP_YES=true ;;
    --overwrite)     BOOTSTRAP_OVERWRITE=true ;;
    --dry-run)       BOOTSTRAP_DRY_RUN=true ;;
    --wipe-scratch)  BOOTSTRAP_WIPE_SCRATCH=true ;;
    --resume-scratch) BOOTSTRAP_RESUME_SCRATCH=true ;;
    --*)             echo "warning: unknown flag: $tok" >&2 ;;
    *)               BOOTSTRAP_OBJECTIVE="${BOOTSTRAP_OBJECTIVE:+$BOOTSTRAP_OBJECTIVE }$tok" ;;
  esac
done
set +f

Run checks in order. If any returns non-zero, respect the contract (hard abort or prompt):

check_no_nazgul_dir || exit $?

check_scratch_state; scratch_rc=$?
case $scratch_rc in
  0) ;;
  12)
    if [ "$BOOTSTRAP_RESUME_SCRATCH" = "true" ]; then
      # Keep the existing scratch tree and continue into Phase 2.
      # Downstream steps detect and reuse pre-existing outputs where present.
      true
    elif [ "$BOOTSTRAP_WIPE_SCRATCH" = "true" ]; then
      rm -rf ./.bootstrap-scratch
    elif [ "$BOOTSTRAP_YES" = "true" ]; then
      echo "error: scratch exists, --yes in effect, cannot prompt; pass --wipe-scratch or --resume-scratch" >&2
      exit 12
    else
      # Interactive: prompt the user and RE-INVOKE the skill with the matching
      # flag. The bash fragment exits here so the LLM can ask the question and
      # restart, instead of the resume path being silently unreachable.
      #
      # LLM: use AskUserQuestion with these options:
      #   header: "Scratch"
      #   question: "A previous bootstrap run left ./.bootstrap-scratch/. How would you like to proceed?"
      #   options:
      #     - "Resume" — "Continue from where the last run stopped"
      #     - "Wipe and restart" — "Delete scratch and start fresh"
      #     - "Abort" — "Stop and keep everything as-is"
      #   Then re-invoke:
      #     - Resume        → /nazgul:bootstrap-project --resume-scratch ...
      #     - Wipe/restart  → /nazgul:bootstrap-project --wipe-scratch ...
      #     - Abort         → stop
      echo "./.bootstrap-scratch/ exists from a prior run." >&2
      exit 12
    fi
    ;;
  *) exit "$scratch_rc" ;;
esac

check_docs_agents_empty; docs_rc=$?
case $docs_rc in
  0) ;;
  11)
    if [ "$BOOTSTRAP_OVERWRITE" = "true" ]; then
      # Actually clear the managed targets so relocation produces a clean
      # bundle instead of silently merging with stale files. Only remove
      # paths this skill owns — never the whole ./.claude/ tree (which may
      # hold unrelated user config).
      rm -rf ./docs ./.claude/agents
      rm -f ./.claude/design-tokens.json ./.claude/design-system.md
    elif [ "$BOOTSTRAP_YES" = "true" ]; then
      exit 11
    else
      # LLM: use AskUserQuestion with these options:
      #   header: "Overwrite"
      #   question: "Non-empty ./docs/ or ./.claude/agents/ detected. What would you like to do?"
      #   options:
      #     - "Overwrite" — "Clear managed targets (./docs/, ./.claude/agents/) and proceed"
      #     - "Abort" — "Stop and keep everything as-is"
      #   If Overwrite: re-invoke with --overwrite flag
      #   If Abort: stop
      echo "Non-empty ./docs/ or ./.claude/agents/ detected." >&2
      exit 11
    fi
    ;;
  *) exit "$docs_rc" ;;
esac

check_git_clean
if [ -n "${BOOTSTRAP_GIT_WARNING:-}" ]; then
  echo "$BOOTSTRAP_GIT_WARNING" >&2
fi

Phase 2 — Objective collection

Create the scratch tree:

mkdir -p ./.bootstrap-scratch/context ./.bootstrap-scratch/docs ./.bootstrap-scratch/agents ./.bootstrap-scratch/.claude

Detect whether this is an existing codebase or an empty project:

detect_project_type
echo "Detected: $BOOTSTRAP_PROJECT_TYPE ($BOOTSTRAP_SOURCE_COUNT source files)"

Determine the objective source using a three-tier priority:

1. If $BOOTSTRAP_OBJECTIVE (parsed in Phase 1) is non-empty, use it regardless of project type. Write a minimal project-spec:

   if [ -n "$BOOTSTRAP_OBJECTIVE" ]; then
     cat > ./.bootstrap-scratch/context/project-spec.md <<SPEC
   # Project Specification
   
   ## Source
   - Method: argument
   - Created at: $(date -u +%Y-%m-%dT%H:%M:%SZ)
   
   ## Vision
   $BOOTSTRAP_OBJECTIVE
   SPEC
   fi
   

2. If brownfield ($BOOTSTRAP_PROJECT_TYPE = "brownfield") and no explicit objective, skip interactive questions entirely. The codebase IS the spec — Discovery will scan it and derive everything. Write a codebase-derived project-spec:

   if [ -z "$BOOTSTRAP_OBJECTIVE" ] && [ "$BOOTSTRAP_PROJECT_TYPE" = "brownfield" ]; then
     PROJECT_NAME=$(basename "$(pwd)")
     cat > ./.bootstrap-scratch/context/project-spec.md <<SPEC
   # Project Specification
   
   ## Source
   - Method: brownfield-auto (derived from existing codebase)
   - Created at: $(date -u +%Y-%m-%dT%H:%M:%SZ)
   - Source files detected: $BOOTSTRAP_SOURCE_COUNT
   
   ## Vision
   Document and analyze the existing $PROJECT_NAME codebase.
   
   ## Note
   This is a brownfield project. The objective, architecture, features, and
   constraints will be derived from scanning the existing source code during
   the Discovery phase. No interactive input was required.
   SPEC
   fi
   

3. If greenfield ($BOOTSTRAP_PROJECT_TYPE = "greenfield") and no explicit objective:

   First, enforce the --yes non-interactive contract — we cannot ask questions in that mode:

   if [ -z "$BOOTSTRAP_OBJECTIVE" ] && [ "$BOOTSTRAP_PROJECT_TYPE" = "greenfield" ] && [ "$BOOTSTRAP_YES" = "true" ]; then
     echo "error: greenfield project with no objective; --yes is set so we cannot prompt." >&2
     echo "       Re-run with an explicit objective argument, e.g.:" >&2
     echo "       /nazgul:bootstrap-project --yes \"your product idea here\"" >&2
     exit 13
   fi
   

   Otherwise, run the condensed Tier 1 interactive flow. Ask these 5 questions one at a time, phrased naturally, and wait for each answer:

   1. "In a sentence or two, what are you building?"
   2. "Who will use this?"
   3. "What are the 3-5 core features?"
   4. "What problem does this solve?"
   5. "Any hard constraints? (Skip if none.)"

   After collecting answers, write ./.bootstrap-scratch/context/project-spec.md with the standard sections (## Vision, ## Target Users, ## Core Features, ## Problem Statement, ## Constraints).

   After Tier 1, offer Tier 2: "Got the basics. Want to go deeper on user stories and success metrics? (~5 more minutes) (y/n)"

   • If yes, ask per-feature user stories and success metrics, append to the spec.
   • If no, finalize with Tier 1 content only.

Verify the file exists before proceeding:

test -f ./.bootstrap-scratch/context/project-spec.md || { echo "error: project-spec.md not written" >&2; exit 30; }

Phase 3 — Pipeline execution

Export STATE_ROOT and BUNDLE_MODE for all downstream renders:

export STATE_ROOT="./.bootstrap-scratch"
export BUNDLE_MODE="true"
source "$CLAUDE_PLUGIN_ROOT/scripts/lib/bootstrap-render.sh"

Render and invoke each pipeline agent. For each agent, the renderer produces a Nazgul-free prompt pointed at the scratch tree; the LLM then executes the agent's instructions.

3a. Discovery

Render the discovery prompt and invoke it:

render_agent_prompt "$CLAUDE_PLUGIN_ROOT/agents/discovery.md" "$STATE_ROOT" > "$STATE_ROOT/.discovery-prompt.md"

Then invoke the discovery agent via the Agent tool, passing the rendered prompt as the system/initial message. The agent writes to $STATE_ROOT/context/{project-profile,project-classification,architecture-map,existing-docs}.md.

After the agent returns, verify expected outputs exist:

for f in project-profile.md project-classification.md architecture-map.md; do
  test -f "$STATE_ROOT/context/$f" || {
    echo "error: discovery did not produce $f; preserving scratch for debugging" >&2
    exit 40
  }
done

3b. Doc generation

render_agent_prompt "$CLAUDE_PLUGIN_ROOT/agents/doc-generator.md" "$STATE_ROOT" > "$STATE_ROOT/.docgen-prompt.md"

Invoke the doc-generator agent. It reads $STATE_ROOT/context/ and writes to $STATE_ROOT/docs/{PRD,TRD,ADR-*,test-plan,manifest}.md (and migration-plan.md if classification=migration).

Verify:

for f in PRD.md TRD.md test-plan.md; do
  test -f "$STATE_ROOT/docs/$f" || {
    echo "error: doc-generator did not produce $f; preserving scratch" >&2
    exit 41
  }
done

3c. Reviewer instantiation

The reviewer template at agents/templates/reviewer-base.md is rendered once per reviewer domain determined from project-profile.md. Available domain definitions live in agents/templates/reviewer-domains.json (existing structure: keys like qa-reviewer, api-reviewer, with per-domain checklist and review_steps arrays, plus description, title, context_items, category, approved_criteria, rejected_criteria). The JSON does NOT have a keywords field today, so domain selection goes via a two-step rule:

1. Always include: qa-reviewer and code-reviewer (baseline; assumed present in the JSON — if a key is missing, skip it with a warning and continue).
2. Conditionally include based on keyword scan of project-profile.md:
   • api-reviewer — if profile mentions api, rest, graphql, endpoint
   • frontend-reviewer — if profile mentions react, vue, svelte, angular, next.js, nuxt, nextjs
   • security-reviewer — always include when the profile mentions auth, login, password, token, jwt, oauth
   • performance-reviewer — if profile mentions database, caching, redis, perf

This mapping is hard-coded in select_reviewer_domains (see Step 2 below). Domains not defined in reviewer-domains.json are skipped.

For each selected domain, render the template with BUNDLE_MODE=true:

DOMAINS=$(select_reviewer_domains "$STATE_ROOT/context/project-profile.md" "$CLAUDE_PLUGIN_ROOT/agents/templates/reviewer-domains.json")

for domain in $DOMAINS; do
  out="$STATE_ROOT/agents/${domain}.md"
  BUNDLE_MODE=true render_template "$CLAUDE_PLUGIN_ROOT/agents/templates/reviewer-base.md" \
    | substitute_domain_vars "$domain" "$CLAUDE_PLUGIN_ROOT/agents/templates/reviewer-domains.json" \
    > "$out"
  test -s "$out" || { echo "error: reviewer $domain rendered empty" >&2; exit 42; }
done

3d. Designer (conditional)

if grep -qiE 'react|vue|svelte|angular|swiftui|flutter' "$STATE_ROOT/context/project-profile.md"; then
  render_agent_prompt "$CLAUDE_PLUGIN_ROOT/agents/designer.md" "$STATE_ROOT" > "$STATE_ROOT/.designer-prompt.md"
  # Invoke designer agent. Writes to $STATE_ROOT/.claude/{design-tokens.json,design-system.md}.
fi

Phase 4 — Transform, relocate, cleanup

Source the relocate helpers:

source "$CLAUDE_PLUGIN_ROOT/scripts/lib/bootstrap-relocate.sh"

Run the transform:

bash "$CLAUDE_PLUGIN_ROOT/scripts/bootstrap-transform.sh" "$STATE_ROOT"
transform_rc=$?
case $transform_rc in
  0) ;;
  2) echo "error: transform usage error" >&2; exit 50 ;;
  3)
    echo "error: transform final assertion failed — see output above." >&2
    echo "       scratch preserved at $STATE_ROOT for debugging." >&2
    exit 51
    ;;
  *) echo "error: transform failed (exit $transform_rc)" >&2; exit 52 ;;
esac

If --dry-run was set (parsed in Phase 1), stop here. Report the scratch location to the user:

if [ "$BOOTSTRAP_DRY_RUN" = "true" ]; then
  echo "Dry-run complete. Review the bundle at $STATE_ROOT/ before re-running without --dry-run."
  exit 0
fi

Relocate atomically:

relocate_bundle "$STATE_ROOT" "." || exit $?

Append to .gitignore and clean up:

append_gitignore "."
cleanup_scratch "$STATE_ROOT"

Phase 5 — Summary

count_md()   { find "$1" -maxdepth 1 -type f -name '*.md' 2>/dev/null | wc -l | tr -d ' '; }
count_dir()  { find "$1" -maxdepth 1 -type f 2>/dev/null | wc -l | tr -d ' '; }

DOCS=$(count_md ./docs)
CTX=$(count_md ./docs/context)
AGENTS=$(count_md ./.claude/agents)
DESIGN=$(count_dir ./.claude)

cat <<SUMMARY
Bootstrap complete.

Generated:
  ./docs/            $DOCS documents (PRD, TRD, ADRs, test plan)
  ./docs/context/    $CTX context files
  ./.claude/agents/  $AGENTS reviewer agents
  ./.claude/         $DESIGN design-system files (if UI surface detected)

Next steps:
  - Review ./docs/PRD.md and ./docs/TRD.md
  - Commit the bundle:      git add docs/ .claude/ && git commit
  - Use the reviewers:      invoke them from Claude Code in this repo
SUMMARY