agent-starter

An opinionated multi-agent skill pack for Claude Code, Codex, and Cursor. Deep, handwritten skills for HCI usability modeling, Apple HIG Doctor guidance, copywriting, code cleanup, and TOON token savings.

No orchestration framework. No aspirational YAML. Just agent-native project files generated from one shared skill source.

Star History

What you get

29 shipped skills:

Skill	Covers
human-processor-model	Estimates task time, cognitive load, memory burden, and perceptual/motor bottlenecks in product flows.
goms-klm-analysis	Decomposes workflows into GOMS/KLM goals, operators, methods, selection rules, waits, and interaction-cost comparisons.
hig-doctor-audit	Runs the HIG Doctor verification loop with `npx hig-doctor`, severity gates, exported reports, and category-to-skill routing.
hig-project-context	Creates shared Apple design context so HIG skills tailor guidance without repetitive questions.
hig-foundations	Apple HIG foundations: color, typography, SF Symbols, dark mode, accessibility, layout, materials, motion, privacy, writing.
hig-platforms	Platform-specific HIG guidance for iOS, iPadOS, macOS, tvOS, visionOS, watchOS, and games.
hig-patterns	Apple UX patterns for onboarding, launch, loading, permissions, feedback, undo, settings, sharing, and collaboration.
hig-inputs	Apple input guidance for gestures, keyboards, pointers, Apple Pencil, Digital Crown, focus, remotes, and spatial input.
hig-technologies	Apple technology guidance for Siri, Apple Pay, HealthKit, ARKit, iCloud, Sign in with Apple, SharePlay, Wallet, and more.
hig-components-content	HIG content display components: charts, collections, image views, web views, color wells, lockups, share sheets.
hig-components-layout	HIG layout and organization: sidebars, split views, tab bars, scroll views, windows, panels, lists, and tables.
hig-components-menus	HIG menus and actions: buttons, context menus, toolbars, menu bar, pop-up buttons, and disclosure controls.
hig-components-search	HIG search and navigation components: search fields, page controls, and path controls.
hig-components-dialogs	HIG dialogs and overlays: alerts, action sheets, popovers, sheets, and digit entry views.
hig-components-controls	HIG controls: pickers, toggles, sliders, steppers, segmented controls, text fields, labels, and validation.
hig-components-status	HIG status and progress UI: progress indicators, status bars, loading states, and activity rings.
hig-components-system	HIG system experiences: widgets, Live Activities, notifications, complications, App Clips, shortcuts, and watch faces.
copywriting-frameworks	Headlines, landing pages, ads, emails, CTAs, AIDA, objections, proof placeholders, critiques.
finish-setup	Provisions a freshly scaffolded SaaS project through the wired MCPs: Stripe products, database migrations, email DNS, analytics, GitHub.
toon-formatter	When TOON helps, when it does not, and how to invoke the TOON commands.
cleanup-all	Orchestrates the full cleanup pipeline.
cleanup-unused	Detects and removes high-confidence dead code, exports, files, and dependencies.
cleanup-cycles	Detects and untangles circular dependencies.
cleanup-dedupe	Extracts high-confidence duplicate code into shared utilities.
cleanup-types	Consolidates duplicated or fragmented type definitions.
cleanup-weak-types	Replaces weak types with stronger inferred or validated types.
cleanup-defensive	Removes pointless try/catch blocks and guards that hide errors.
cleanup-legacy	Removes zero-caller deprecated, legacy, and fallback paths.
cleanup-slop	Removes unhelpful narration comments while preserving useful WHY comments.

Agent Targets

Agent	Generated output	Notes
Claude Code	`.claude/`	Native Claude skills, settings, TOON slash commands, optional hooks.
Codex	`AGENTS.md` + `.codex/skills/*/SKILL.md`	Root Codex guidance points to project-local skill files.
Cursor	`.cursor/rules/*.mdc`	Cursor project rules generated as Agent Requested rules, plus an always-applied skill-selection rule.

Claude remains the default for backwards compatibility. Use --agent all to install all supported targets.

Install

First run / project setup:

# Claude Code only (default)
npx create-agent-starter@latest

# Codex only
npx create-agent-starter@latest --agent codex

# Cursor only
npx create-agent-starter@latest --agent cursor

# Claude Code + Codex + Cursor
npx create-agent-starter@latest --agent all

# Backwards-compatible aliases still work
npx create-claude-starter@3.0.1 --agent all

Optional global CLI for repeated sync, status, and add commands:

npm i -g create-agent-starter

agent-starter sync
agent-starter status
agent-starter add mcp neon
agent-starter add skill cleanup-types

For team setup, prefer the npx create-agent-starter@latest ... commands above so contributors do not need a preinstalled global binary.

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

npm i @toon-format/toon gpt-tokenizer

Profiles

npx create-agent-starter@latest --profile all --agent all
npx create-agent-starter@latest --profile apple-hig --agent codex,cursor
npx create-agent-starter@latest --profile design-hci --agent codex
npx create-agent-starter@latest --skills copywriting-frameworks,cleanup-unused --agent cursor

Profiles select a skill set. Agent targets decide where that skill set is installed.

Stack profiles (next-saas, next, node, base) additionally bundle MCP servers for the stack. init auto-detects the right one from package.json.

agent.json

init writes an agent.json manifest at the project root — the single declaration of the project's agent environment (profile, targets, skills, MCP servers). Check it into git; every contributor then runs:

npx create-agent-starter@latest sync

or agent-starter sync after installing the global CLI, and gets identical native config for whichever agent they use: skills plus .mcp.json (Claude Code), .codex/config.toml + AGENTS.md (Codex), and .cursor/mcp.json + rules (Cursor). Sync is idempotent — generated sections are fenced with markers, manual edits outside them survive, and MCP entries agent-starter didn't write are never touched.

npx create-agent-starter@latest status            # diff agent.json vs native configs; exits 1 on drift
npx create-agent-starter@latest add mcp neon      # catalog: github, neon, stripe, resend
npx create-agent-starter@latest add mcp internal --command ./bin/mcp --env API_KEY='${API_KEY}'
npx create-agent-starter@latest add skill cleanup-types

Secrets are referenced as ${VAR} and resolved by each agent at runtime — never written to the generated files. Sync appends missing vars to .env.example and warns when they're unset.

The next-saas profile is the flagship: cleanup + copywriting skills, the finish-setup provisioning skill, and Neon/Stripe/Resend/GitHub MCPs — designed as the companion to a Next.js SaaS template. After scaffolding, open your agent and say "finish setup": the agent creates Stripe products to match your billing plans, verifies migrations, and walks email DNS via the wired MCPs.

Structure

.claude/
  skills/<skill>/skill.md
  commands/
  utils/toon/cli.mjs

.codex/
  config.example.toml
  skills/<skill>/SKILL.md
AGENTS.md

.cursor/
  rules/agent-starter.mdc
  rules/<skill>.mdc
  rules/<skill>/references/

The package keeps one shared source of truth in templates/.claude/skills/ and generates Codex/Cursor formats from that source during install.

The Codex target includes .codex/config.example.toml as a safe project-local example. Keep machine trust, live ~/.codex state, auth, sessions, SQLite databases, and plugin caches out of product repos. See docs/codex-machine-policy.md and docs/untrusted-repo-intake.md.

The Apple HIG skills are vendored from HIG Doctor, including the progressive-disclosure references/ corpus with canonical Apple source links and attribution. The hig-doctor-audit skill points agents at HIG Doctor's published audit CLI and verification workflow.

Benchmarks

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

bun run bench:generate # deterministically regenerate sample workloads
bun run bench          # writes bench/RESULTS.md and bench/results.json

The Markdown file is for humans; bench/results.json is the machine-readable artifact for diffs, CI uploads, and downstream analysis. bun run bench also enforces a conservative 40% aggregate savings gate by default; override it with BENCH_MIN_SAVINGS_PCT=... or --min-savings-pct ... when intentionally changing the workload mix.

Requirements

Node.js >= 18
Claude Code, Codex, or Cursor, depending on the selected target
Optional: @toon-format/toon and gpt-tokenizer for Claude TOON slash commands

License

MIT. Not affiliated with Anthropic, Apple, OpenAI, Cursor, or @toon-format/toon.

Claude Starter