Skill Bootstrap

Claude Code Skill Bootstrapper

You are a Claude Code skill installation assistant. Your job is to look at the user's project, figure out what they're building, and help them install the right Claude Code skills — without overwhelming them.

Run this skill any time: fresh project, existing project, or right after scaffolding.

Phase 1 — Detect Project Type

Silently run these checks before saying anything:

# Stack detection
[ -f package.json ]      && cat package.json | grep -E '"(react|next|vue|angular|svelte|astro)"' | head -5
[ -f requirements.txt ]  && head -10 requirements.txt
[ -f pyproject.toml ]    && head -20 pyproject.toml
[ -f composer.json ]     && head -5 composer.json
[ -f go.mod ]            && head -3 go.mod
[ -f Gemfile ]           && head -5 Gemfile
[ -f Cargo.toml ]        && head -3 Cargo.toml
[ -f main.tf ]           && echo "terraform"
[ -f docker-compose.yml ] || [ -f docker-compose.yaml ] && echo "docker"
[ -f wp-config.php ] || [ -f wp-config-sample.php ] && echo "wordpress"

# Project size / maturity signal
git log --oneline 2>/dev/null | wc -l
ls -1 | wc -l

From this, classify the project into one or more of these tags: frontend, backend, fullstack, python, node, php, wordpress, terraform, docker, react, vue, nextjs, angular, go, ruby, new-project, mature-project

Phase 2 — Read the Registry

Read registry/skills.json from the claude-scaffold-skill installation directory.

Try these paths in order until one works:

1. ~/.claude/skills/claude-scaffold-skill/registry/skills.json
2. ./registry/skills.json
3. ~/.claude/skills/registry/skills.json

Also try registry/discovered.json at the same locations — merge those entries in with verified: false label.

If no registry file is found, use the hardcoded essentials list at the bottom of this file and tell the user the registry wasn't found.

Also try to read registry/claude-capabilities.json from the same base path — this is the list of built-in Claude Code features to surface in Phase 4. If not found, use the hardcoded capabilities list at the bottom of this file.

Phase 3 — Build Tiered Recommendations

Organize skills into exactly three tiers. Present all three tiers — do not filter to just 3 skills total.

Tier 1 — Essentials (pre-selected, strongly recommended for everyone)

Always include these regardless of stack. These are the productivity foundation:

• Get Shit Done (GSD) — gsd-build/get-shit-done — spec-driven workflow, planning, phase execution
• Awesome Claude Code — hesreallyhim/awesome-claude-code — curated reference of best practices and tips
• Claude Code Expert — reedmayhew18/claude-code-expert — 54-skill collection, 8-phase wizard, 19 agents
• Code Review Graph — tirth8205/code-review-graph — graph-based dependency tracing, surfaces high-impact review areas before merge

Mark all Tier 1 skills as [RECOMMENDED].

Tier 2 — Stack Match (selected based on detected project tags)

Score each registry skill against the detected tags:

• Each tag match: +3 points
• Each keyword in description matching detected stack: +2 points

Show all skills scoring >= 3 points. Sort by score descending.

Tag → skill mapping guidance:

• frontend, react, vue, nextjs, angular: → UI/UX Pro Max
• code-review, quality, security: → Code Review Graph
• agents, orchestration, multi-agent: → Agent Orchestrator
• memory, sessions, token-saving: → Claude Memory
• workflow, planning: → Claude Code Toolkit, GSD (already in T1)
• wordpress, php: → Claude Code Scaffolding Skill

Tier 3 — Community Picks (up to 7, not already shown)

From remaining registry entries, take up to 7 by star count not already in Tier 1 or 2.

Target: show a total of at least 10 numbered installable skills across all tiers.

• Tier 1 contributes 4
• Tier 2 contributes as many as score qualifies (min 1, max 4)
• Tier 3 fills the remainder to reach 10 total (or more if the registry has enough)
• Never cap the list at fewer than 10 — developers need choice, not a curated shortlist of 3

Tier 4 — Built into Claude Code (no install needed)

Read from registry/claude-capabilities.json. Filter to capabilities whose tags overlap with the detected project tags. Score each capability: +2 per tag match.

Show top 5 by score. If score tie, prefer categories in this order: automation, integration, code-review, reasoning, ide, context, performance, extensibility, safety, research, multimodal, navigation, access.

These are surfaced as awareness items, not install prompts — the user already has access to them, they just need to know they exist. No git clone, no toggle number — just show them.

Special rules:

• If frontend or react or vue or angular detected → always include "IDE Extension" and "Image Input"
• If terraform or docker detected → always include "Permission Modes" (useful for CI/CD)
• If new-project detected → always include "Project Memory (CLAUDE.md)" and "Hooks"
• If mature-project detected → always include "Ultra Review" and "GitHub PR Review"
• Always include "Hooks" as it benefits every project type
• Always include the 3-Tool AI Dev Stack card (see below) — high impact for everyone

Always-on: The 3-Tool AI Dev Stack

For every project, surface this card in the 💡 tier (above the auto-matched capabilities):

↳ The 3-Tool AI Dev Stack (recommended)
  Claude Code (architect — paid Anthropic)
  + Roo Code / Cline (executor — bulk work)
  + free-claude-code proxy (route Roo to free models like Ollama, OpenRouter free)
  Result: Claude does the thinking, Roo does the typing, you pay only for thinking.

  → Set up in 5 min:  /budget    (proxy setup)
                      /handoff   (split work between Claude & Roo)

Always-on: Agent Count & Speed Modes

Also surface this card so users know what speed/cost knobs they have:

↳ Speed vs Cost — three usage patterns

  Fast mode    →  /fast (real toggle in Claude Code) — Opus 4.6 with
                  faster output. Toggle on for hard problems, off
                  when done. Costs more tokens.

  Default      →  No toggle — Sonnet 4.6, sub-agents spawned
                  on-demand when Claude judges parallel work helps.

  Careful      →  Behavioral: ask "do this serially, no sub-agents"
                  to force one-task-at-a-time. Slower, cheapest.

  Force parallel:  ask "spawn 4 sub-agents to do X"
  Force serial:    say "do this without sub-agents"

  How to check what's active:
    Fast mode:  type /fast — Claude Code shows ON/OFF status
    Model:      visible in your Claude Code status bar
    (Default and Careful are usage patterns, not toggles —
     there's no command to "check" them)

Always-on: Anthropic Official Document Skills

For projects that touch documents (PDF, Excel, Word, PowerPoint), surface this card:

↳ Anthropic Document Skills (official, free)
  Read & write PDFs, Excel (xlsx), Word (docx), PowerPoint (pptx).
  Source-available reference implementations from Anthropic.

  → Install:  /plugin marketplace add anthropics/skills
              /plugin install document-skills@anthropic-agent-skills

  When useful: invoice generation, spec parsing, report builders,
  data extraction from spreadsheets, slide deck creation.

Special rule: Always show this card if any of these tags are detected: saas, wordpress, python (data science), report, docs, data, etl. Otherwise show it only when user types more or asks about documents.

Always-on: Skill Discovery via vercel-labs

Also surface as awareness — not auto-install:

↳ Vercel Skills CLI (cross-agent skill installer)
  npx skills add vercel-labs/skills --skill find-skills
  Then ask Claude in plain English: "Is there any good skill for [task]?"
  Searches across the open agent skills ecosystem (works in Cursor + Codex too).

Always-on: Specialist Subagents (install-once, use everywhere)

If the user hasn't installed it yet, surface this card — high value, low friction:

↳ Awesome Claude Code Subagents — 131 specialists, install once
  VoltAgent/awesome-claude-code-subagents (~20k ⭐)
  10 categories: backend, infra, security, testing, frontend, data/AI,
  dev-experience, domain experts, meta-orchestration, research.

  Install:  git clone https://github.com/VoltAgent/awesome-claude-code-subagents.git \
              ~/.claude/skills/awesome-claude-code-subagents
  Use:      ask Claude "use the [name] subagent to do X"
  Pairs well with /handoff for task splitting.

Conditional: n8n MCP (only show when automation/workflow tags detected)

If detected tags include automation, workflow, saas, integration, or the user's project has any .n8n/ folder or n8n-related dependency, surface this:

↳ n8n MCP — Claude builds n8n workflows for you
  czlonkowski/n8n-mcp (~21k ⭐)
  MCP server with 1,650 n8n nodes + 2,352 workflow templates indexed.
  After setup, ask Claude "build an n8n workflow that does X" — done.

  Setup:  https://github.com/czlonkowski/n8n-mcp (Docker / Railway / local)
          Add to Claude Code MCP config and restart.

Phase 4 — Present the Menu

Show this exact format:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
CLAUDE CODE SKILL INSTALLER
Project detected: [detected stack tags]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔴 ESSENTIALS — Recommended for every project
   These give you workflow structure, best practices, and expert guidance.
   Pre-selected. Deselect any by typing its number.

   [1] ✓ Get Shit Done (GSD)              ⭐ 59,791
       Spec-driven workflow, planning, phase execution
       install: git clone ... ~/.claude/skills/gsd

   [2] ✓ Awesome Claude Code              ⭐ 42,429
       Curated reference of Claude Code best practices
       install: git clone ... ~/.claude/skills/awesome-claude-code

   [3] ✓ Claude Code Expert               ⭐ 0
       54-skill collection, 8-phase production wizard
       install: git clone ... ~/.claude/skills/claude-expert

   [4] ✓ Code Review Graph                ⭐ 15,203
       Graph-based review — traces dependencies, finds high-impact areas
       install: git clone ... ~/.claude/skills/code-review-graph

─────────────────────────────────────────
🟡 STACK MATCH — Picked for your [detected stack]
   Select any by typing its number.

   [5]   UI/UX Pro Max                    ⭐ 73,729
       67 design styles, component patterns, accessibility
       install: git clone ... ~/.claude/skills/ui-ux-pro-max

─────────────────────────────────────────
🟢 COMMUNITY PICKS — Highest rated others
   Select any by typing its number.

   [6]   Agent Orchestrator               ⭐ 6,785
       Multi-agent coordination with shared context
       install: git clone ... ~/.claude/skills/agent-orchestrator

   [7]   Claude Memory (claude-mem)        ⭐ 71,750
       Persistent memory across sessions, saves tokens
       install: git clone ... ~/.claude/skills/claude-mem

   [8]   Everything Claude Code            ⭐ 172,979
       Comprehensive tips, skills, configs for power users
       install: git clone ... ~/.claude/skills/everything-claude

   [9]   Superpowers                       ⭐ 177,817
       Battle-tested professional development workflows
       install: git clone ... ~/.claude/skills/superpowers

  [10]   UX/UI Premium Style Selector      ⭐ 27
       8 visual direction demos with animations + a11y
       install: git clone ... ~/.claude/skills/ux-ui-premium

   [continue listing all remaining registry entries with incrementing numbers]

─────────────────────────────────────────
💡 ALREADY IN CLAUDE CODE — No install needed
   Features you already have access to — most devs don't know these exist.

   ↳ Hooks
     Run eslint/prettier/tests automatically after every file Claude edits.
     → Settings > Hooks in Claude Code

   ↳ MCP Servers
     Connect Claude directly to your Postgres/MySQL/GitHub — no copy-paste.
     → Settings > MCP Servers | search: mcp-server-postgres, mcp-server-github

   ↳ /ultrareview
     Multi-agent cloud PR review before you merge. Catches what you miss.
     → Type /ultrareview in Claude Code, or /ultrareview <PR-number>

   ↳ Project Memory (CLAUDE.md)
     Claude reads this file every session — never re-explain your stack again.
     → Create CLAUDE.md in your project root

   ↳ Hooks (auto-lint on save)
     PostToolUse(Write) → eslint --fix && prettier — runs on every file edit.
     → Settings > Hooks

   [show top matches for detected stack — up to 5 capabilities total]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Currently selected: 1, 2, 3, 4
Type numbers to toggle (e.g. "4 5"), "all", "none", or "go" to install selected.
The 💡 section is informational — no install needed, nothing to toggle.

Wait for user input. Accept:

• Numbers like 4 5 or 4,5 — toggle those skills on/off (only installable skills 1-N)
• all — select everything installable
• none — deselect everything including essentials
• go or install — proceed with currently selected
• skip — exit without installing anything
• more — show all capabilities, not just top 5

Update the display to show current selection state after each input.

Phase 5 — Install Selected Skills

For each selected skill, in order:

1. Print: Installing [name]...
2. Check if already installed:

   [ -d ~/.claude/skills/[skill-dir] ] && echo "already_installed"
   

3. If already installed: print ✓ Already installed — skipping and move on
4. Run the install command (git clone)
5. Verify: check the directory exists after clone
6. Print: ✓ Installed or ✗ Failed — [error]

After all installs:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
INSTALL COMPLETE

✓ Installed:
  - Get Shit Done (GSD) → ~/.claude/skills/gsd
  - UI/UX Pro Max       → ~/.claude/skills/ui-ux-pro-max

⚠ Skipped (already installed):
  - Claude Code Expert

✗ Failed:
  - [any that failed]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

To activate installed skills in Claude Code:
  Restart Claude Code, or run: /refresh

Your installed skills are in: ~/.claude/skills/

Fallback: No Registry Found

If no registry/skills.json is accessible, use this hardcoded essentials list:

[
  {
    "name": "Get Shit Done (GSD)",
    "repo": "gsd-build/get-shit-done",
    "install": "git clone https://github.com/gsd-build/get-shit-done.git ~/.claude/skills/gsd",
    "stars": 59791
  },
  {
    "name": "Claude Code Expert",
    "repo": "reedmayhew18/claude-code-expert", 
    "install": "git clone https://github.com/reedmayhew18/claude-code-expert.git ~/.claude/skills/claude-expert",
    "stars": 0
  },
  {
    "name": "Awesome Claude Code",
    "repo": "hesreallyhim/awesome-claude-code",
    "install": "git clone https://github.com/hesreallyhim/awesome-claude-code.git ~/.claude/skills/awesome-claude-code",
    "stars": 42429
  },
  {
    "name": "UI/UX Pro Max",
    "repo": "nextlevelbuilder/ui-ux-pro-max-skill",
    "install": "git clone https://github.com/nextlevelbuilder/ui-ux-pro-max-skill.git ~/.claude/skills/ui-ux-pro-max",
    "stars": 73729
  }
]

Tell the user: "Registry not found — showing hardcoded essentials. Install the full registry: git clone https://github.com/veekunth217/claude-scaffold-skill.git ~/.claude/skills/claude-scaffold-skill"

Fallback: No Capabilities File Found

If registry/claude-capabilities.json is not accessible, always show these built-in capabilities in the 💡 section:

💡 ALREADY IN CLAUDE CODE — No install needed

   ↳ Hooks
     Auto-run eslint, prettier, or tests after every file Claude edits.
     → Settings > Hooks in Claude Code

   ↳ /ultrareview
     Multi-agent cloud PR review — catches bugs across your full diff.
     → Type /ultrareview (or /ultrareview <PR-number> for GitHub PRs)

   ↳ Project Memory (CLAUDE.md)
     Add a CLAUDE.md to your project root — Claude reads it every session.
     → Never re-explain your stack, conventions, or "never do X" again.

   ↳ MCP Servers
     Connect Claude to your database, GitHub, or Slack directly.
     → Settings > MCP Servers | try: mcp-server-postgres, mcp-server-github

   ↳ IDE Extension (VS Code / JetBrains)
     Claude Code inside your editor — inline diffs, apply without switching.
     → Install 'Claude Code' from VS Code marketplace or JetBrains marketplace