메뉴
Use when initializing the workflow system for a project — bootstraps structure, analyzes the codebase, generates CLAUDE.md with adaptive philosophy, skills, and rules. Re-run to find drift, gaps, and stale configuration.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when you want to sweep every built-but-unsigned-off work record — or recap and sign off on ad hoc work from this same conversation that has no work record at all — and give each one a human verdict, approve or request changes. The durable acceptance gate distinct from tests passing (/test) and code-quality review (/review). Keywords - acceptance, sign-off, demo, verification brief, human verdict, demo:pending, session-recall.
Use when you need a quick reference for available commands, want to see workflow status, or need a recommendation for what to do next.
Use when the backlog needs hygiene — review stale backlog records, parked-trigger wakes, unsynced local records, and orphaned plans/worktrees
Use when /claude-tweaks:review passes and you need to capture learnings, clean up specs/plans, update skills, and decide next steps. The lifecycle closure step.
Use when you need to create, update, query, or resolve open items in a pipeline ledger file. Called by /claude-tweaks:build, /claude-tweaks:test, /claude-tweaks:review, /claude-tweaks:wrap-up, and /claude-tweaks:flow — or standalone for ledger inspection.
Use when implementing a work record, spec, or design doc end-to-end. Accepts a record reference (#N) or spec number (legacy alias) for full lifecycle tracking, or a design doc path to skip /claude-tweaks:specify and build directly from brainstorming output.
| name | claude-tweaks:init |
| description | Use when initializing the workflow system for a project — bootstraps structure, analyzes the codebase, generates CLAUDE.md with adaptive philosophy, skills, and rules. Re-run to find drift, gaps, and stale configuration. |
Interaction style: Present single decisions via the
AskUserQuestiontool (options with one marked Recommended) instead of a plain-text numbered list. For multi-item decisions, render a batch table with recommended actions pre-filled, then capture the apply-all/override decision via oneAskUserQuestioncall. Never make more than oneAskUserQuestioncall per logical decision — resolve each before showing the next. End skills with a## Next Actionsblock rendered viaAskUserQuestion(context-specific options, one recommended), not a navigation menu.
Bootstrap the workflow system for a project AND generate intelligent configuration from codebase analysis. Handles everything from directory creation to CLAUDE.md generation, skills, rules, and journey discovery — in one command.
[ /claude-tweaks:init ] → /claude-tweaks:capture → /claude-tweaks:challenge → /superpowers:brainstorming → /claude-tweaks:specify → /claude-tweaks:build → /claude-tweaks:stories → /claude-tweaks:test → /claude-tweaks:review → /claude-tweaks:wrap-up
^^^^ YOU ARE HERE ^^^^
This skill works for both greenfield and brownfield projects, operating in two modes:
| Mode | Trigger | Behavior |
|---|---|---|
| Initial | No .claude/ directory or CLAUDE.md | Full bootstrap + reconnaissance → generate everything from scratch |
| Update | Existing .claude/ config found | Skip bootstrap (idempotent) + diff-based audit → propose targeted patches |
/claude-tweaks:help or /claude-tweaks:build fails because something is missingIf $ARGUMENTS is provided, treat it as:
~/projects/their-app) — cd there first--update or update — force Update mode even if the config looks minimalbootstrap — run Phase 0 only (structure + deps), then stopconfig — run Phases 0 + 2 + 3 + 5 (bootstrap + recon + CLAUDE.md)skills — run Phases 0 + 2 + 3 + 4 + 6 (bootstrap + recon + skills)journeys — run Phases 0 + 8 (bootstrap + journey discovery)docs — run Phases 0 + 2 + 3 + 8.5 (bootstrap + doc registry)Every scope above still runs Phase 9 as its terminal summary/confirm/write step, except bootstrap (which stops the invocation after Phase 0) — this includes the goal-based scopes (config, skills, journeys, docs) even though none of them list Phase 9 explicitly in their phase subset above. The interactive Scope Selection Gate's own early-stop choices (Option 4 "Done," and Option 2 Interactive's per-phase "Done") are the other paths that stop before Phase 9; see "Finalizing the worktree.always Decision" for why this distinction matters.
If no arguments, analyze the current working directory. Phase 0 runs first, then a scope selection gate determines which remaining phases to run (see "Scope Selection Gate" below).
| Phase | What Happens | Output |
|---|---|---|
| 0 | Bootstrap structure (dirs, files, deps, git, browser) | Workflow infrastructure ready |
| 1 | Determine mode (Initial vs Update) | Mode decision + existing config inventory |
| 2 | Codebase reconnaissance (8 detection steps) | Raw findings: stack, architecture, conventions, pain points, maturity |
| 3 | Build profile (Initial) or drift report (Update) | Stack Profile or Configuration Health Report |
| 4 | Generate skill manifest | Scored + prioritized skill candidates. Priority 2-3 → backlog |
| 5 | Generate / update CLAUDE.md | CLAUDE.md (how to work here). Improvement pain points → backlog |
| 6 | Generate / update skills | SKILL.md files for approved skills. Aspirational skills → backlog |
| 7 | Generate / update rules (optional) | Path-scoped .claude/rules/ files |
| 8 | Discover user journeys (optional) | Journey files or skeleton backlog work records |
| 8.5 | Create doc registry | docs/REGISTRY.md. Doc work → backlog |
| 9 | Present summary and confirm | Final confirmation before writing files |
Fast, idempotent structural setup. Creates directories, starter files, and verifies dependencies. Skips anything that already exists.
Core Bootstrap (Steps 1–8):
Verify Superpowers plugin and the built-in code-simplifier subagent are available. Read bootstrap-steps.md (Step 1) for the dependency list, detection commands, and install hint.
Create the required workflow directories — idempotent, only creates what's missing. Read bootstrap-steps.md (Step 2) for the directory taxonomy and per-directory rationale.
Create specs/INDEX.md — only if missing, never overwrite. Read bootstrap-steps.md (Step 3) for the canonical starter content.
Suggest entries for transient workflow artifacts; never modify .gitignore without asking. Read bootstrap-steps.md (Step 4) for the full suggested block and the stories-commit prompt.
Confirm the directory is a git repo; warn if not (review and wrap-up will be degraded). Read bootstrap-steps.md (Step 5) for the full procedure.
Ensure .worktrees/ exists in the project root for the git-fallback path; leave any .claude/worktrees/ directory alone as a separate, harness-owned convention that needs no migration. Also offers the worktree.always policy opt-in (recommended default: on) — the decision is queued here but the file write is deferred to avoid this same run denying its own later writes; see "Finalizing the worktree.always Decision" and "Worktree Policy Finalization" below. Read bootstrap-steps.md (Step 6) for the full procedure.
Detect agent-browser; surface the install command if missing. Never block init, never auto-install, never prompt for backend choice. Read bootstrap-steps.md (Step 7) for the full procedure.
Detect Node (and optionally git), install the statusline wrapper at ~/.claude-tweaks/bin/statusline.js, and prompt before wiring statusLine.command in ~/.claude/settings.json — never overwrite a non-claude-tweaks command. Read bootstrap-steps.md (Step 8) for the full procedure (detection, package-manager prompts, settings.json migration matrix, NO_COLOR opt-out).
Optional Enhancements (Steps 9–15):
GitHub issue form template offer (agent-task.yml). Read bootstrap-steps.md (Step 9) for the full procedure.
When Phase 2 detects frontend signals, present the three-option Impeccable setup prompt (Full / Plugin-only / Skip) and write the design-integration flag to CLAUDE.md — the /claude-tweaks:design wrapper reads this as Layer 1 of its detection logic. Read bootstrap-steps.md (Step 10) for the full procedure (frontend-detection list, install sequence, flag-value table, re-run behavior, failure handling).
Always offered (not frontend-gated). Present the two-option diagram-suggestions prompt (Enable / Skip) and write the diagram-suggestions flag to CLAUDE.md under the existing ## Design integration section. Soft-hook nudges in /journeys, /specify, and /review read this flag to decide whether to suggest invoking /claude-tweaks:visualize. No install step — /claude-tweaks:visualize is a native skill. Read bootstrap-steps.md (Step 11) for the full procedure.
When frontend signals are detected and components.json doesn't exist (or exists without full AI-agent wiring), present the shadcn/ui setup prompt (Full / CLI-only / Skip, or the narrower "wire remaining layers" offer when the CLI is already initialized) and write the shadcn-integration flag to CLAUDE.md. Currently write-only — no other skill reads the flag yet. Read bootstrap-steps.md (Step 12) for the full procedure (framework/package-manager detection, install sequence, MCP/skills wiring, flag-value table, re-run behavior, failure handling).
Always offered (not gated) — detect which claude-tweaks skills ship a routine-template.yml (plus any named routine-template-<variant>.yml siblings) without an existing instantiated record for this project, present them via one multiSelect AskUserQuestion call (grouped into ≤4-option questions when there are more than 4 candidates) with their default schedules, and invoke /claude-tweaks:routine create <skill> [--variant=<name>] --defaults --environment=<id> --source init for each selected candidate — no per-candidate interactive walkthrough. Idempotent: candidates with an existing record are never re-offered. Read bootstrap-steps.md (Step 13) for the full procedure.
Offer only on projects with a GitHub remote — writes .github/workflows/track-issue-fixes.yml, which labels (fix-on-<branch>) and comments on issues fixed on non-default branches, then strips those labels once the fix reaches the default branch and GitHub closes the issue natively. Idempotent: skipped silently once the workflow file exists. Read bootstrap-steps.md (Step 14) for the full procedure.
Decide whether the unified work record — used by /claude-tweaks:capture, /claude-tweaks:specify, /claude-tweaks:triage, /claude-tweaks:dispatch, /claude-tweaks:tidy, and the health skills — is backed by GitHub issues or local record files, and write work-backend to CLAUDE.md under a ## Work records section (gated on the same GHE-safe two-tier remote check Step 9 uses). Then probe GitHub-native capabilities once (work-types, work-links) and offer to provision all 20 core labels now. See _shared/work-record.md for the taxonomy these config keys govern. Read bootstrap-steps.md (Step 15) for the full procedure.
If Step 6 (bootstrap-steps.md) queued a worktree.always decision, it must be written to .claude-tweaks/policy.yml exactly once, as the very last filesystem action before this /init invocation ends — for whatever reason it ends. Phase 9's "Worktree Policy Finalization" (below) is the normal place this happens: per "Input" above, every scope reaches Phase 9 except bootstrap, including the goal-based scopes (config/skills/journeys/docs) even though none of them list Phase 9 in their own phase subset. The known early-exit paths that stop the invocation before Phase 9 ever runs are: $ARGUMENTS was bootstrap (stops after Phase 0); the Scope Selection Gate's Option 4 ("Done"); or Option 2 (Interactive)'s own per-phase gate, if the user selects "Done" ("Stop here") after any phase. These are the known cases, not necessarily an exhaustive list of every way this invocation could ever end — whatever the actual reason this invocation is ending, write the decision right there, immediately before it ends: create .claude-tweaks/ if it doesn't exist, then write or update the worktree.always: line in .claude-tweaks/policy.yml (merge into existing content — preserve every other line in the file untouched; create the file with just that one line if it didn't exist). If the decision was "Yes," tell the user: "worktree.always is now enforced — your next edit requires an isolated worktree; run /superpowers:using-git-worktrees first."
If this invocation instead reaches Phase 9, the decision is finalized there — see "Worktree Policy Finalization."
After Phase 0 completes, present the scope selection — unless $ARGUMENTS already specified a goal-based scope (e.g., bootstrap, config, skills, journeys, docs), in which case skip this gate and run the corresponding phases.
Not silenced by auto. The scope-selection gate is on the "What auto does NOT silence" list in _shared/auto-mode-contract.md — it is a project-shape governance decision that requires explicit user input regardless of auto state. The prompt below always renders unless $ARGUMENTS already specified a scope.
Call AskUserQuestion (see "Phases at a Glance" above for the full table; Phase 8 is auto-marked "Skip — no UI detected" when reconnaissance finds no user-facing surface):
question: "Bootstrap complete. How much setup do you want?", header: "Setup scope", multiSelect: falselabel: "Auto (Recommended)", description: "Run all included phases without stopping"label: "Interactive", description: "Pause for confirmation between phases"label: "Essentials", description: "Reconnaissance + CLAUDE.md only (phases 2, 3, 5)"label: "Done", description: "Just needed the bootstrap structure"Option 1 (Auto): Run all included phases end-to-end. Phase 3 auto-confirms classification when detection confidence is high and signals are consistent (otherwise presents the confirmation gate as a KEPT-PROMPT). Phase 4 still presents the skill selection (governance decision — never silenceable). Phase 9 still presents the final summary for confirmation (governance decision). All other phases run without pausing.
Option 2 (Interactive): After each phase completes, present its output, then call AskUserQuestion. This is a template — re-issue it once per phase (not a single static site), substituting {N}, {N+1}, {description}, and {N+2} each time:
question: "Phase {N} complete. Continue to Phase {N+1} ({description})?", header: "Phase gate", multiSelect: falselabel: "Continue (Recommended)", description: "Proceed to Phase {N+1} ({description})"label: "Skip Phase {N+1}", description: "Move to Phase {N+2}"label: "Done", description: "Stop here"If the user selects this template's "Done" and Step 6 queued a worktree.always decision, write it now — see "Finalizing the worktree.always Decision" above.
Option 3 (Essentials): Runs phases 2, 3, 5 only. Produces CLAUDE.md with proper philosophy and Don'ts. Defers skills, rules, journeys, and doc registry for later (suggest re-running /init or using goal-based arguments).
Option 4 (Done): Stop after Phase 0. The user has the directory structure, starter files, and dependencies — they'll configure manually or run /init again later. If Step 6 queued a worktree.always decision, write it now — see "Finalizing the worktree.always Decision" above.
When a phase is excluded (by interactive skip, essentials mode, or goal-based argument), handle its dependents:
| If skipped | Impact | Handling |
|---|---|---|
| Phase 2 (recon) | Phases 3-8.5 lose their input | Skip all dependent phases — cannot generate config without reconnaissance |
| Phase 3 (profile) | Phases 4, 5, 8.5 lose maturity/tier classification | Skip dependent phases — philosophy and doc tier need classification |
| Phase 4 (manifest) | Phase 6 has no skill list | Skip Phase 6 |
| Phase 5 (CLAUDE.md) | No downstream dependency | Safe to skip |
| Phase 6 (skills) | No downstream dependency | Safe to skip |
| Phase 7 (rules) | No downstream dependency | Safe to skip |
| Phase 8 (journeys) | No downstream dependency | Safe to skip |
| Phase 8.5 (doc registry) | No downstream dependency | Safe to skip |
When skipping a phase due to a missing dependency, note it: "Skipping Phase {N} ({name}) — requires Phase {dep} which was excluded."
Before any deep analysis, check what already exists:
Check:
- CLAUDE.md exists? Read it.
- .claude/skills/ directory? List all skills, read each SKILL.md frontmatter.
- .claude/rules/ directory? List all rules, read each.
- .claude/settings.json? Read it.
Update Mode runs three sub-phases before deciding whether to continue with the full reconnaissance:
Update Mode procedures live in update-mode.md — load only when Phase 1 detects existing config. That file contains the inventory template, the contract-drift marker table, and the early-exit decision logic.
Work through these detection steps systematically. Use parallel tool calls aggressively — all glob/grep operations within a substep are independent and should run concurrently.
Detect:
- README.md, CONTRIBUTING.md, docs/ — project purpose and domain
- LICENSE — open source vs proprietary
- .github/, .gitlab-ci.yml, Jenkinsfile, .circleci/ — CI/CD platform
- Monorepo vs single app (workspaces config, multiple package.json, apps/, packages/)
- Age — earliest commit date (`git log --reverse --format="%ai" -1`)
- Activity — commits in last 90 days, number of contributors
Steps 2b–2g cover stack detection, architecture detection, convention detection, workflow detection, pain point detection, and existing AI configuration detection. For the complete detection tables and checklists, read detection-tables.md in this skill's directory.
Assess the project's maturity stage (greenfield / pre-launch / early-production / established) to inform the Philosophy section in CLAUDE.md. Gather signals in parallel and classify the project.
See Phase 2h Maturity Detection in detection-tables.md (this skill's directory) for the full signal table and classification criteria. Carry the classification forward to Phase 3, where it is presented alongside the doc tier for unified confirmation. Do NOT present the classification for confirmation here — Phase 3 is the single confirmation gate for all project classifications.
Initial Mode produces a Stack Profile (identity, stack table, architecture, conventions, workflows, health indicators, skill candidates) followed by a unified Project Classification confirmation. Update Mode produces a Drift Report (covered, stale, drifted, gap classifications). Both require user confirmation before proceeding.
For the complete profile and drift report templates, read profile-templates.md in this skill's directory.
After presenting the Stack Profile (or Drift Report), present the unified Project Classification gate. This is the single gate where the user confirms or overrides all downstream decisions about philosophy and doc structure.
Decision logic:
high / med / low) for both maturity and doc-tier dimensions.auto mode is set AND both dimensions are high confidence AND signals are internally consistent. Log to decisions.md and proceed.auto with med/low confidence, or contradictory signals (e.g., greenfield code + production-grade infra).For the confidence rubric, auto-confirm log format, and full confirmation-gate template (table + philosophy + doc structure + numbered options), read phase-3-classification.md in this skill's directory.
Carry the confirmed maturity and doc tier forward to Phase 5 (CLAUDE.md Philosophy) and Phase 8.5 (Doc Registry).
Initial Mode: Score and prioritize all skill candidates.
Update Mode: Score only the gaps. Existing skills that need updating are handled as patches, not new skills.
Parallel execution: Use parallel tool calls aggressively — scoring of independent skill candidates is read-only (re-checking grep/glob signals from Phase 2) and should run concurrently.
Parallel execution (conditional): When the candidate list has ≥ 8 skills, dispatch scoring as parallel Task agents per the Subagent Contract (
_shared/subagent-output-contract.md). Otherwise, run the scoring inline in the main thread.
Apply the Frequency + Complexity + Danger rubric (max 9). Generate skills scoring 6+ first. Skills not selected (Priority 2-3 or aspirational) become backlog work records with their scoring rationale and Phase 2 evidence — no reconnaissance is wasted.
For the full scoring rubric (with examples), the skill-category mapping reference, and the deferred-skill backlog work-record format, read phase-4-scoring.md in this skill's directory. For the Skill Manifest presentation template and the selection prompt, read profile-templates.md (Phase 4 section).
CLAUDE.md describes how to work in this codebase — the patterns to follow, the commands to run, the conventions to respect, and the mistakes to avoid. Every entry should help someone working in the codebase right now. Things that don't exist yet belong in the backlog, not here.
Initial Mode generates CLAUDE.md from scratch with sections for Stack, Structure, Commands, Conventions, Philosophy, Testing, Environment, Git, and Don'ts. Update Mode produces targeted patches, not rewrites. The Philosophy section adapts to detected project maturity. The Don'ts section is the highest-ROI output — mine it from convention conflicts and observed anti-patterns (not from missing infrastructure).
For the complete CLAUDE.md template, patch format, Philosophy generation guide, Don'ts mining guide, and principles, read claude-md-template.md in this skill's directory.
Phase 2f findings split into CLAUDE.md Don'ts (convention conflicts and anti-patterns) and backlog work records (missing infrastructure, practices, stale deps, dead code). For the backlog work-record templates and the "Pain Points Routed" summary template, read pain-point-routing.md in this skill's directory.
Initial Mode generates full SKILL.md files for each approved skill. Update Mode produces targeted patches for drifted skills and full SKILL.md for gap skills. Each generated skill must pass quality gates (codebase-grounded examples, working commands, project-specific anti-patterns). Skill depth scales with complexity score.
Only generate skills for patterns that actually exist and are actively used in the codebase. Aspirational skills (e.g., testing for a project with no tests) should have been captured as backlog work records in Phase 4 — do not generate SKILL.md files for them.
For the complete SKILL.md template and depth guide, read skill-template.md in this skill's directory. For the drift-patch procedure and quality gates applied to drifted/gap skills, read _shared/harness-health-analysis.md — the same procedure /claude-tweaks:wrap-up Step 7 and the standalone /claude-tweaks:harness-health routine use.
Generate .claude/rules/ files for path-specific conventions (e.g., "all files in src/api/ must use the error handler"). Project-wide conventions belong in CLAUDE.md, not rules. For the rule frontmatter template, common rule candidates, and Update-Mode hint, read rules-template.md in this skill's directory.
For projects with user-facing surfaces (web app, CLI, API with docs), offer to discover and document user journeys — especially valuable for brownfield projects with features but no documented flows. Skip for pure libraries or when docs/journeys/ already has comprehensive coverage.
This project has user-facing features but no documented user journeys.
User journeys help /review test the app against experiential expectations.
Would you like to discover and document journeys?
1. Yes — scan codebase for routes and user flows, create journey files **(Recommended)**
2. Yes, with browser — scan codebase AND walk the app in a browser for richer "should feel" details
3. Skip — I'll add journeys later
Use Phase 2 findings to identify routes/pages, infer personas (user roles, auth flows, public vs. authenticated), and group routes into goal-oriented journey skeletons. Write skeleton files to docs/journeys/ (use journey-template.md from the /claude-tweaks:journeys skill directory) — each one stamped **Status:** Skeleton — inferred from code, not yet browser-tested. Skeletons document the intended flows but "should feel" fields are weaker until browser-tested.
Capture enrichment as backlog work records for each skeleton journey:
### Browser-test journey: {name}
Skeleton journey inferred from code — "should feel" and "red flags" need browser validation.
Routes covered: {list}. Persona: {persona}. Dev URL: {if known}.
Run `/claude-tweaks:review journey:{name}` to enrich with experiential details.
Delegate to /claude-tweaks:visual-review discover — runs the full 6-phase discovery process (codebase scan → journey candidates → browser walkthrough → write files → coverage report → handoff). Ask the user for the dev server URL before invoking.
Note that the user skipped journey discovery. Suggest running /claude-tweaks:visual-review discover later when they're ready.
Create the documentation registry that maps project docs to the code areas they cover. This registry is consumed by /build (Step 6.5) to auto-update docs when relevant code changes, and by /wrap-up (Step 6) for final sweep and registry maintenance.
Use the confirmed doc tier from Phase 3 — do not re-detect.
For the complete procedure (registry format, tier definitions, standard folder taxonomy, common Auto-detect patterns, inventory → assess → batch table → create → backlog-capture flow, and Update-Mode diff logic), read docs-structure.md in this skill's directory.
Present a consolidated summary of all work done across Phase 0 (bootstrap) and Phases 1-8 (configuration). Wait for user confirmation before writing generated files. This phase is the terminal step for every scope except bootstrap — see "Input" above — including the goal-based scopes (config/skills/journeys/docs) and the Scope Selection Gate's Option 1 (Auto) and Option 3 (Essentials), even though none of the goal-based scopes list Phase 9 explicitly in their own phase subset.
Both modes lead with a Verified & Consistent section — an affirmative report of what was checked and found healthy (dependencies present, contract markers up to date, config items still accurate, detections confirmed), not just what changed or was created. This is required, not optional: Update Mode in particular often proposes few or no changes, and the user needs to see what was audited and deliberately left alone. The early-exit fast path (Phase 1u.6) carries its own shorter Verified & Consistent block.
For the complete summary templates for both modes, read summary-templates.md in this skill's directory.
After writing files, surface what was created. Generate the table from the actual artifacts produced this run (only include rows for actions that actually occurred):
| Action | Detail | Ref |
|---|---|---|
| Bootstrap | Created specs/, docs/, docs/journeys/, .worktrees/, etc. (only missing dirs) | Step 2 |
| Starter files | Wrote specs/INDEX.md (only if missing) | Step 3 |
| Statusline | Installed wrapper at ~/.claude-tweaks/bin/statusline.js; wired ~/.claude/settings.json | Step 8 |
| Design integration | Set design-integration: {enabled/plugin-only/disabled} in CLAUDE.md | Step 10 |
| shadcn integration | Set shadcn-integration: {enabled/cli-only/disabled} in CLAUDE.md | Step 12 |
| Work records | Set work-backend / work-types / work-links in CLAUDE.md; offer 17-label bootstrap | Step 15 |
| Routines | Instantiated {N} routine(s): {list} (or "Offered, none set up") | Step 13 |
| Worktree policy | Set worktree.always: {true/false} in .claude-tweaks/policy.yml (only if Step 6 asked this run) — written last, after every other row above, to avoid mid-run self-lockout; see "Worktree Policy Finalization" below | Step 6 |
| Classification | Confirmed maturity {value}, doc tier {N} | Phase 3 |
| CLAUDE.md | Wrote {N} lines (Initial) / Applied {N} patches (Update) | Phase 5 |
| Skills | Generated {N} SKILL.md files: {list} | Phase 6 |
| Rules | Created {N} path-scoped rules in .claude/rules/ | Phase 7 |
| Journeys | Wrote {N} skeleton journey files (or delegated to /visual-review discover) | Phase 8 |
| Doc registry | Created docs/REGISTRY.md with {N} entries | Phase 8.5 |
| Backlog | Added {N} work records (deferred skills, pain points, doc work, skeleton enrichment) | Phases 4-8.5 |
Execute only after user confirmation.
Write this AFTER every write in the Actions Performed table above has completed — it must be the very last filesystem action of the entire /init invocation. If Step 6 (bootstrap-steps.md) queued a worktree.always decision, write it now: this is the deferred write described in Step 6, deferred specifically so this run's own Steps 7-14, Phases 1-8.5, and this same Phase 9's own confirmed generated-file writes (the Actions Performed table above) were never blocked by a policy that turned on mid-run. (The bootstrap-only scope already wrote its queued decision immediately after Step 14 — see "Finalizing the worktree.always Decision" after Phase 0 — so there is nothing to do here for that scope.)
Create .claude-tweaks/ if it doesn't exist. Read .claude-tweaks/policy.yml if present; if it has an existing worktree.always: line, replace that line, otherwise append a new worktree.always: {true|false} line (create the file with just that line if it didn't exist). Preserve every other line in the file untouched.
If the decision was "Yes," tell the user: "worktree.always is now enforced — your next edit requires an isolated worktree; run /superpowers:using-git-worktrees first."
Resolve the recommended action from the signals that fired during this run. This lookup table is the assistant's own resolution logic — it stays internal and is never itself shown to the user or converted into an AskUserQuestion option. Resolve signals top-to-bottom; the first matching row is the recommendation. The signal rows are not exhaustive over every possible post-init state (e.g. Update Mode completing normally with zero drift and no backlog writes matches none of them) — when no signal row matches, use the Fallback row so there is always a defined recommendation.
| Signal | Recommended Next Action |
|---|---|
| Update Mode ran AND total drift count > 0 | /claude-tweaks:tidy — clean up drifted/stale config and backlog items before resuming feature work |
| Backlog has work records written this run (deferred skills, pain points, doc work, skeleton enrichment) | /claude-tweaks:tidy — triage what /init just captured |
| Initial Mode ran AND backlog is empty | /claude-tweaks:capture {idea} — capture the first idea or feature into the backlog for triage |
| Everything is clean (Update Mode early-exit OR Initial Mode with nothing routed to the backlog) | /claude-tweaks:help — see the full lifecycle overview and current pipeline status |
| Fallback (no row above matches) | /claude-tweaks:help — see the full lifecycle overview and current pipeline status |
Once resolved to a single recommended row, call AskUserQuestion with exactly 3 options — the resolved recommendation, plus the two "Always" actions below:
question: "What's next?", header: "Next step", multiSelect: falselabel: a short one-line summary of it suffixed (Recommended), description: the full command text from the matched rowlabel: "Specify next feature", description: "/claude-tweaks:specify {first feature topic} — jump straight to specifying the first lifecycle feature"label: "Tidy backlog", description: "/claude-tweaks:tidy — review backlog entries"If the resolved recommendation is itself /claude-tweaks:tidy (rows 1 or 2), it and Option 3 refer to the same command — collapse them into a single (Recommended) option rather than presenting /claude-tweaks:tidy twice, leaving 2 options for that call instead of 3.
| Pattern | Why It Fails |
|---|---|
| Modifying existing backlog work records or INDEX.md content | Phase 0 is additive — it creates missing files but must not overwrite user content |
| Skipping CLAUDE.md generation | Without CLAUDE.md, /claude-tweaks:review can't find verification commands |
| Running init in a non-git directory without warning | /claude-tweaks:review and /claude-tweaks:wrap-up depend on git — the user should know about degraded behavior |
| Installing browser tools without asking | Browser integration is optional — surface the install command but never run npm install automatically |
| Prompting for a browser backend choice | There is only one backend (agent-browser) — do not present a choice |
Silently rewriting a legacy backlog-backend flag to work-backend during Phase 0 | That rename is Update-Mode's job, offered as a staged change (see update-mode.md's Work-Record Backend Drift) — Phase 0 must leave an existing ## Backlog integration section untouched |
Generating generic skills (e.g., auth.md, api-routes.md) | These are not real conventions — they're feature names. Real skills must encode rules, anti-patterns, or "Why this is done this way" insights grounded in patterns actually observed in the codebase. If the project doesn't use WebSockets, don't create a realtime skill. If it has no tests, capture testing as an aspirational backlog item, not a SKILL.md file. |
| Generating generic skills not grounded in the codebase | Skills must encode observed patterns — generic advice adds noise, not value |
| Rewriting CLAUDE.md in Update Mode | Update Mode produces patches, not rewrites — existing config embeds hard-won lessons |
| Over-generating skills (15 mediocre > 5 excellent) | Each skill must earn its existence by encoding knowledge that would otherwise be lost |
| Skipping team input | Code archaeology alone misses social conventions — PR process, deploy cadence, naming debates |
| Aspirational Don'ts for things that don't exist | Don'ts are guardrails for existing patterns, not wishes for missing infrastructure. "No CI" is a backlog item, not a Don't. |
| Putting improvement ideas in CLAUDE.md | CLAUDE.md describes how to work in the codebase as it is — improvement opportunities belong in the backlog with Phase 2 context |
| Generating skills for patterns that don't exist yet | Aspirational skills (testing for a project with no tests) become backlog work records with Phase 2 evidence, not SKILL.md files |
| Hardcoding greenfield philosophy for all projects | The Philosophy section must adapt to detected project maturity — what's correct for a greenfield project is dangerous for an established one |
| Creating doc files with only TODO placeholders | Phase 2 reconnaissance has the data — generate real content grounded in actual findings. If a doc would be < 20 lines of real content, it belongs in README instead of its own file. |
| Skipping journey discovery for projects with user-facing features | Journeys are what /review tests against — without them, visual QA has no experiential anchor |
| Writing journey "should feel" without actually using the app | Codebase-only skeletons are a starting point but the "should feel" is weaker — mark them as skeletons |
--full to force the complete pass.| Skill | Relationship |
|---|---|
/claude-tweaks:capture | First skill to use after /claude-tweaks:init — add ideas to the backlog |
/claude-tweaks:help | Shows workflow status — useful to verify /claude-tweaks:init worked. Surfaces doc staleness signals from the registry. |
/claude-tweaks:review | /review lens 3i uses the doc registry to check documentation freshness. |
/claude-tweaks:visual-review | Phase 8 (hybrid mode) delegates to /visual-review discover for browser-assisted journey discovery. |
/claude-tweaks:build | /init creates docs/REGISTRY.md (Phase 8.5) that /build consumes in Step 6.5 for documentation sync |
/claude-tweaks:wrap-up | Captures learnings after features — keeps generated skills alive and accurate. Step 7 references skill-template.md from /claude-tweaks:init's directory. /wrap-up Step 6 maintains the doc registry created by /init. |
/claude-tweaks:tidy | /tidy Step 4.6 audits doc registry health — flags stale entries, gaps, pattern drift. Suggests /init update for tier drift. |
/claude-tweaks:browse | Depends on agent-browser, which /claude-tweaks:init detects (and surfaces install instructions for) in Phase 0 |
/claude-tweaks:design | Step 10 sets up Impeccable design integration (install plugin + CLI, optionally run init) and writes the design-integration kill-switch flag to CLAUDE.md that the wrapper reads as Layer 1 of its detection logic. |
/claude-tweaks:visualize | Step 11 offers to enable diagram suggestions and writes the diagram-suggestions flag to CLAUDE.md — no install step, this skill is native. Soft-hook nudges in /journeys, /specify, and /review read the flag to decide whether to suggest invoking it. |
shadcn/ui (companion) | Step 12 offers to bootstrap the shadcn CLI, wire its official MCP server into .mcp.json, and install shadcn's official Skill (skills add shadcn/ui) for live Claude Code project context. Writes the shadcn-integration flag to CLAUDE.md — currently write-only, no other skill reads it yet. |
/superpowers:using-git-worktrees | /claude-tweaks:init optionally configures the worktree directory that using-git-worktrees needs |
/claude-tweaks:stories | /init Phase 8 (journey discovery) feeds /stories — discovered journeys become input for story generation |
/claude-tweaks:version | /version reads the same plugin.json that /init may print during bootstrap |
/claude-tweaks:routine | Step 13 discovers claude-tweaks skills shipping a routine-template.yml (plus any named routine-template-<variant>.yml siblings) with no existing instantiated record, presents them via one multiSelect AskUserQuestion call (grouped into ≤4-option questions when there are more than 4 candidates) with their default schedules, resolves environment once, then invokes /claude-tweaks:routine create <skill> [--variant=<name>] --defaults --environment=<id> --source init for each selected candidate — pure discovery + handoff, no logic duplicated. |
/claude-tweaks:harness-health and _shared/harness-health-analysis.md | Phase 6's drift-patch procedure and Phase 3/1u's skill classification apply this shared procedure instead of an inline copy, sharing its judgment logic and .claude-tweaks/harness-health/ cursor/cache state with /claude-tweaks:wrap-up Step 7 and the standalone routine. |
_shared/work-record.md | Step 15 provisions the work-backend / work-types / work-links config keys this file documents as the record taxonomy's driver and capability contract. The label-provisioning offer (Step 15c) runs _shared/label-bootstrap.md's canonical label list, which this file names as the taxonomy home. |
_shared/auto-mode-contract.md | Single source of truth for auto-mode behavior — read before adding any auto-mode handling. Phase 3 classification auto-confirm follows the contract's confidence-gated pattern. |
| All workflow skills | Depend on the structure /claude-tweaks:init creates |