claude-starter

An opinionated Claude Code skill pack for fintech devs and Anthropic power-users. Six deep, handwritten skills plus a thin CLI for TOON — a JSON compression format that typically cuts input tokens 40–60% on tabular data.

No orchestration framework. No aspirational YAML. Just skills that activate when you need them.

What you get

6 top-level skills — auto-activate on keywords:

5 slash commands for TOON:

• /convert-to-toon <file> — encode + report measured savings
• /analyze-tokens <file> — compare JSON vs TOON token counts without writing a file
• /toon-encode <file>, /toon-decode <file>, /toon-validate <file>

All TOON commands shell out to the canonical @toon-format/toon npm library via a 90-line wrapper at .claude/utils/toon/cli.mjs, and use gpt-tokenizer for real token counts (OpenAI BPE — directionally accurate proxy for Claude; for exact counts use Anthropic's /v1/messages/count_tokens endpoint).

Install

# Into current project
npx create-claude-starter@latest

# Into a specific dir, no prompts
npx create-claude-starter@latest ./my-app --yes

For TOON commands, add the runtime deps to your project:

npm i @toon-format/toon gpt-tokenizer

Use

Skills auto-activate on context:

User: How do I verify a Stripe webhook signature?
Claude: [stripe-expert activates] Use stripe.webhooks.constructEvent...

User: Convert api-response.json to TOON
Claude: [runs /convert-to-toon api-response.json]
  ✓ Wrote api-response.toon
  Tokenizer: gpt-tokenizer (OpenAI BPE — approximate proxy for Claude)
  JSON:      4,587 tokens (12,840 bytes)
  TOON:      2,759 tokens (7,128 bytes)
  Saved:     1,828 tokens (39.8%)

Why this exists

The Claude Code skill market is fragmented: Anthropic ships 17 general-purpose skills and nothing for fintech or platform integrations. Community mega-packs (awesome-claude-code-toolkit, antigravity-awesome-skills) compete on volume — 1,400+ skills with razor-thin depth.

This pack goes the other way: 6 skills, hand-maintained, each genuinely better than what's out there. The Stripe skill alone is 2,100+ lines of tested integration patterns. If you're building a fintech app or extending Claude itself, this is the starting point.

Not in this repo (and why)

• Orchestration / semantic matching / multi-skill workflows — removed. Claude Code selects skills natively via frontmatter description; the previous TypeScript orchestration engine was placeholder code.
• YAML workflow engine — removed. Out of scope for a skill pack.
• Meta-commands / command registry / validators — removed. Claude Code's built-in /skill and plugin system handle this.
• Native Zig TOON binary — removed. The canonical @toon-format/toon npm package (1.8M downloads/month) ships pure JS, cross-platform.
• Blockchain / iOS / Shopify / Whop skills — removed. Niche or thin; if you need them, spin a focused pack.
• Duplicated code-quality skills — removed. Anthropic already ships cleanup-unused, cleanup-slop, etc. natively.

Structure

.claude/
├── skills/
│   ├── stripe/        # 2,100+ lines
│   ├── supabase/
│   ├── plaid/         # consolidated: Link + Auth + Transactions + Identity + Accounts
│   ├── expo/          # consolidated: core + EAS Build + EAS Update + Expo Router
│   ├── anthropic/     # 1 main skill + 6 Claude Code meta-tooling sub-skills
│   └── toon-formatter/
├── commands/
│   ├── convert-to-toon.md
│   ├── analyze-tokens.md
│   ├── toon-{encode,decode,validate}.md
│   └── {discover,install}-skills.md
└── utils/toon/cli.mjs # 90-line wrapper around @toon-format/toon + gpt-tokenizer

Benchmarks

Real measured token counts for a handful of representative workloads are in bench/. Numbers use gpt-tokenizer, not a claimed heuristic.

Requirements

• Node.js ≥ 18
• Claude Code ≥ 1.0
• (Optional) @toon-format/toon and gpt-tokenizer in your project for TOON commands

License

MIT. Not affiliated with Stripe, Supabase, Plaid, Expo, or Anthropic.