customize（customize 搜索）

Customize Setup for Your Project

Customize this Claude Code setup for: $ARGUMENTS

Critical

• Do NOT modify hooks, agents, or other skills — this skill ONLY fills CUSTOMIZE markers
• Do NOT invent information — if the user didn't provide something, ask for it
• Keep it concise — CLAUDE.md is loaded on every prompt. Verbose sections waste tokens.
• Match the style of existing content in each file when filling markers

What This Skill Does

This setup has <!-- CUSTOMIZE --> markers in CLAUDE.md and .claude/rules/*.md that need project-specific details. This skill collects your project information ONCE, then fills in ALL markers automatically.

See examples.md in this skill directory for a complete example of what good customization looks like (the "Acme HR" reference project).

Task Tracking

Tasks survive context compacts — skipping this check causes lost progress and repeated work.

Before starting work, run TaskList to check if tasks already exist from a previous session or before a compact. If tasks exist:

1. Read existing tasks with TaskGet for each task ID
2. Find the first task with status pending or in_progress
3. Resume from that task — do NOT recreate the task list

If no tasks exist, create them after the first round of user questions:

Example task list:

Task 1: Collect project brief (all rounds)
Task 2: Read all CUSTOMIZE markers
Task 3: Fill CLAUDE.md
Task 4: Fill rule files
Task 5: Clean up markers
Task 6: Validate completeness

Mark each task in_progress when starting and completed when done.

Step 1: Collect Project Brief

Use AskUserQuestion to collect project details. Ask up to 4 questions at a time, then follow up with remaining questions. Do NOT proceed until all critical information is gathered.

Round 1 — Core Identity

Ask these questions using AskUserQuestion:

1. Project name and description — "What is your project name and a one-line description?"

- Example: "Acme SaaS — workforce management platform"

1. App structure — "Is this a monorepo or single app?"

- If monorepo: ask for app names, ports, and purposes in follow-up - If single-app: note this and skip monorepo section

1. Package manager — "Which package manager?"

- Options: npm, pnpm, yarn, bun

1. Component library — "What UI component library do you use?"

- Options: shadcn/ui, Radix UI, MUI, custom, none - Follow up: "What is the import path?" (e.g., @/components/ui, ~/components)

Round 2 — Architecture

1. Auth model — "What is your auth/account model?"

- Options: Single-user, Multi-tenant with teams, Multi-tenant with orgs, Other - If multi-tenant: "How are accounts scoped? (account_id FK, org_id, workspace_id)"

1. Framework wrappers — "Does your framework provide wrappers for Server Actions or auth?"

- Example: "enhanceAction from MakerKit", "withAuth from next-auth" - If no wrapper: we'll use the default Zod + manual auth pattern

1. Logging — "What logging approach do you use?"

- Options: console (dev only), Pino, Winston, framework logger - Follow up if not console: "What is the import path?" (e.g., @/lib/logger)

1. Supabase client paths — "Where are your Supabase client factories?"

- Default: @/lib/supabase/server and @/lib/supabase/client - Or framework-provided: @kit/supabase/server-client

Round 3 — Commands & Git

1. Key commands — "List your main dev commands:"

- dev, build, test, lint, verify/check-all, migration commands, type generation - If unsure, ask to read package.json scripts

1. Git strategy — "What is your branch strategy?"

- Options: main only, main + development, feature branches to main - Remote names (origin, upstream if forked from a template)

1. CI/CD — "What CI/CD do you use?"

- Options: GitHub Actions, Vercel, GitLab CI, none yet - Key jobs: typecheck, lint, test, deploy

Round 4 — Optional Features

1. i18n — "Do you use internationalization?"

- If yes: library (next-intl, react-i18next), translation namespaces

1. Testing structure — "Where do your tests live?"

- Default: __tests__/{feature}/ - Or: src/__tests__/, app/**/*.test.ts, colocated

1. Anything else — "Any framework-specific patterns, admin panel requirements, feature flags, or special conventions I should know about?"

Step 2: Read All CUSTOMIZE Markers

Before filling anything in, read every file that has markers to understand context:

CLAUDE.md
.claude/rules/git-workflow.md
.claude/rules/database.md
.claude/rules/patterns.md
.claude/rules/coding-style.md
.claude/rules/security.md
.claude/rules/ui-components.md
.claude/rules/forms.md
.claude/rules/testing.md
.claude/rules/i18n.md
.claude/rules/admin.md
.claude/rules/pages-and-layouts.md
.claude/rules/route-handlers.md

Read each file completely before editing.

Step 3: Fill CLAUDE.md

This is the most important file — Claude reads it on every conversation.

Marker: Project description (top of file)

Replace the <!-- CUSTOMIZE: Replace with a brief description --> comment with a one-line description. Example: Acme SaaS — Next.js App Router, Supabase, TypeScript.

Remove the entire HTML comment block (the template instructions at the top of the file).

Marker: Critical Rules

Add any framework-specific critical rules from the intake. Examples:

• If using a Server Action wrapper: "Skipping {wrapper} means unauthenticated data can reach the database."
• If using a specific logger: "Use getLogger() from {path} instead of console.log."
• If monorepo with upstream: "When merging upstream, propagate infrastructure changes to all product apps."

Marker: Monorepo

• If monorepo: Add a table with app names, ports, and purposes
• If single-app: Remove the ## Monorepo section entirely

Marker: Commands

Add all commands from the intake in a code block. Group by category (dev, test, build, deploy).

Marker: Architecture

Describe: multi-tenant model, data fetching patterns, auth, RLS approach, any special systems.

Marker: Verification

Add the project's verify/check commands (typecheck, lint, test).

Step 4: Fill Rule Files

For each rule file, replace <!-- CUSTOMIZE --> markers with project-specific content.

git-workflow.md

• Remotes table (origin URL, upstream if applicable)
• Branch strategy (adjust the table)
• Verify command
• CI pipeline table
• Upstream merge process (if forked from template, otherwise remove section)

database.md

• Migration/typegen command paths
• RLS helper functions (list any existing helpers, or keep the example)
• OTP components (if framework provides them, otherwise remove marker)

patterns.md

• Server Action wrapper (replace manual pattern with framework wrapper if applicable)
• Supabase client import paths (update the table)
• Auth/account model (adjust or remove multi-tenant section)

coding-style.md

• Server Action wrapper reference (match patterns.md)

security.md

• Auth wrapper documentation

ui-components.md

• Component library name and import path

forms.md

• Server Action wrapper
• Form component library imports

testing.md

• Test directory structure
• Test commands

i18n.md

• i18n setup (library, config)
• Translation namespaces
• If no i18n: Delete the entire file (or leave as-is for future use)

admin.md

• Admin verification pattern

pages-and-layouts.md

• Loading component
• Workspace providers
• Feature flags (if applicable)

route-handlers.md

• Route handler wrapper (if framework provides one)

Step 5: Clean Up Markers

After filling all markers:

1. Remove all <!-- CUSTOMIZE --> comments — they served as prompts and should be deleted after customization. The content they prompted for is now filled in.
2. Remove the template instruction block at the top of CLAUDE.md (the HTML comment with "SETUP INSTRUCTIONS").
3. Update file headers — Change generic titles to project-specific ones:

- # Patterns — Next.js Supabase TypeScript could become # Patterns — Acme Next.js Supabase

Step 6: Validate

Run these checks to confirm completeness:

Check 1: No remaining CUSTOMIZE markers

grep -rn "CUSTOMIZE" CLAUDE.md .claude/rules/ 2>/dev/null | grep -v "node_modules" || echo "All markers filled!"

If any remain, go back and fill them. If a marker cannot be filled due to missing information, ask the user.

Check 2: CLAUDE.md has content in all sections

Read back CLAUDE.md and verify:

• Project description is present (not a placeholder)
• Critical Rules section has content
• Monorepo section has content OR was removed for single-app
• Commands section has a code block with real commands
• Architecture section describes the actual architecture
• Verification section has real commands

Check 3: No template instruction block remains

grep -c "SETUP INSTRUCTIONS" CLAUDE.md || echo "Template header removed!"

Report

Print a summary:

Customization complete for {project-name}!

Files modified:
  - CLAUDE.md
  - .claude/rules/git-workflow.md
  - .claude/rules/database.md
  - ... (list all modified files)

Remaining markers: {count} (0 = fully customized)

Next steps:
  1. Review the changes: git diff
  2. Test the setup: start a new Claude Code session and verify context
  3. Optional: Add project-specific validators to post_tool_use.py

Resuming After Context Compact

If you notice context was compacted or you're unsure of current progress:

1. Run TaskList to see all tasks and their status
2. Find the in_progress task — that's where you were
3. Run TaskGet {id} on that task to read full details
4. Continue from that task — don't restart from the beginning

Tasks persist across compacts. The task list is your source of truth for progress, not your memory.

Pattern for every work session:

TaskList → find in_progress or first pending → TaskGet → continue work → TaskUpdate (completed) → next task

Troubleshooting

Markers remain after Step 6 validation

The most common cause is a marker inside an HTML comment that looks like content. Read the file, find the <!-- CUSTOMIZE comment, and either fill it or delete it. If the marker needs information you don't have, use AskUserQuestion to ask.

User unsure about commands or paths

Offer to read their package.json for commands: Read package.json and extract scripts. For monorepo structures, Glob "apps/*/package.json" reveals app layout.

i18n section not applicable

If the project doesn't use i18n, you can either delete .claude/rules/i18n.md entirely or leave it with a note at the top: "i18n is not currently used in this project."

User provides minimal answers

Ask follow-up questions. A one-word answer like "pnpm" is fine for package manager, but "yes" for monorepo needs follow-up: "What are your app names, ports, and purposes?"