sneg55/agent-starter/new-project
Claude Code and agent skill by sneg55/agent-starter
- community
Checking that the SKILL.md file is still in the repository.
Full project bootstrap - interviews the developer (name, description, stack, components), then scaffolds directory structure, CLAUDE.md, config files, hooks, skills, and first commit.
Install
npx vskill@latest install sneg55/agent-starter/new-projectSKILL.md
New Project Bootstrap
<!-- Mirrors AGENT.md in the agent-starter repo. If guides change, update this skill to match. -->
Use when starting a new project from scratch. Scaffolds a complete AI-friendly project following the agent-starter patterns: feature-based directory structure, CLAUDE.md with memory taxonomy, config files, optional hooks and skills, first commit.
For existing projects, use /adopt-project instead.
Phase 0: Detect what's already installed
Hooks and skills install system-wide under ~/.claude/, so they're shared across every project. Detect them first and never ask about components that are already present. Run:
Interpret the output:
- Hooks installed if
.agent-starter-versionexists (or hook scripts are
found). Record the stamped version.
- Skills installed per directory listed as
present.
Carry this into the interview and scaffold: only ask about, and only install, what's missing. If a stamped hooks version is present but older than the repo VERSION (checked in Phase 1 once you have the repo path), note that an update is available and offer to re-run install.sh (idempotent) - don't force it.
Phase 1: Interview
Ask these questions one at a time before taking any action:
- Project name - what is the name of the project?
- Description - one sentence describing what it does.
- Tech stack - language, framework, package manager (e.g. "TypeScript, Next.js, pnpm").
- Tailwind design-system lint - ask this follow-up only when the stack is
TypeScript/JavaScript with a UI framework (React, Next.js, Remix, Vite + React): "Does the project use Tailwind v4, and should I add @shadcn/lint (blocks raw palette colors, arbitrary values, inline styles, unknown classes, and restyling design-system components via className)? yes/no". Record the answer; it drives the optional block in Phase 2 step 3. Skip the question for non-UI stacks.
- Optional components - ask only about what Phase 0 reported as missing.
If hooks and all skills are already installed, skip this question entirely - state what was detected ("Hooks v0.4.4 and all 8 skills already installed system-wide, skipping") and move on. Otherwise offer the missing set:
- Hooks (auto-enforce file size limits, lint-on-save, silent-error and dangerous-command blocking, codebase health checks at
~/.claude/hooks/) - Skills (commit, commit-push-pr, simplify, remember, dream, new-project, adopt-project, reflect at
~/.claude/skills/) - Both
- Neither
- Repo path - what is the local path to the agent-starter repo? (e.g.
~/code/agent-starter). Always required: the CLAUDE.md template, foundation templates, and lint configs are all copied from the repo. (Hooks and skills also install from here when selected and not already present.)
Do not proceed past this step until you have all answers.
Phase 2: Scaffold
Execute these steps in order.
1. Create directory structure
Design principle: organize by feature, not by technical layer. Each feature gets its own directory under src/features/ with ALL related files (implementation, types, constants, validation, tests). Keep files under 200 lines each. Shared type definitions go in src/types/ to break import cycles. Named constants go in src/constants/ (no magic strings anywhere).
2. Generate CLAUDE.md
Copy the canonical template - do not inline or hand-write it, so it never drifts from templates/CLAUDE.md (which carries the Memory System, Git Safety, Implementation Notes, and Self-improvement loop sections):
Then fill in the ## Project-Specific Instructions section at the bottom:
3. Create config files
`.gitignore` at <project-name>/.gitignore:
`.env.example` at <project-name>/.env.example:
`README.md` at <project-name>/README.md:
`CLAUDE.local.md` at <project-name>/CLAUDE.local.md (gitignored above - personal, machine-local instructions that never get committed):
`.claude/rules/` - modular instruction files the agent loads alongside CLAUDE.md. Create the directory and the apply-on-touch pattern index, which is the same file /adopt-project writes (Tier 4):
Write <project-name>/.claude/rules/starter-patterns.md:
Optionally also create topic stubs (testing.md, git-workflow.md, code-style.md, security.md) per templates/NEW_PROJECT_PROMPT.md - offer these but don't force them.
Lint configs - if the stack is TypeScript/JavaScript:
Tailwind design-system lint (opt-in, only if the Phase 1 follow-up was yes):
Then wire the fragment into eslint.config.mjs: add import shadcnRules from './eslint.shadcn.mjs' next to the other imports and ...shadcnRules, as the last entry of the tseslint.config(...) call. Set settings.shadcn.ui in the fragment to the project's component import prefix (@/components/ui is the shadcn default) and the **/components/ui/** override to the directory that owns the components. @shadcn/lint needs ESLint 9.30+ and Tailwind v4; it finds the theme without a components.json, so shadcn/ui itself is not required.
If the stack is Python:
See guides/lint-rules-for-ai.md for what the rules catch. Skip for other stacks.
4. Copy foundation templates (TypeScript/JavaScript or Python)
Reference: guides/error-id-registry.md, guides/zod-at-the-boundary.md, guides/large-codebase-best-practices.md
These are the "create from day one" foundation files (see templates/NEW_PROJECT_PROMPT.md -> Foundation Files): a centralized env boundary, a numbered error registry, and an output truncator. Copy them for the matching stack and adapt import paths to the layout.
If the stack is TypeScript/JavaScript:
If the stack is Python:
Skip for other stacks (Rust, Go, etc.) - point the developer at the guides above to build equivalents.
5. Install hooks (if selected and not already installed)
Skip if Phase 0 detected hooks already installed and the stamped version matches the repo VERSION - they're system-wide, so they already cover this project. If installed but stale, offer to update by re-running the installer (it's idempotent). Otherwise run it now.
The installer copies the hooks (and lib/) to ~/.claude/hooks/, stamps the installed version, and merges the hook wiring into ~/.claude/settings.json with jq. Existing entries are preserved and re-running never duplicates anything - do not hand-edit the JSON:
Hook behavior (wired by default):
check-file-size.sh- runs after every Write/Edit. Blocks (exit 2) files over 300 lines; warns over 200 lines. Skips.md,.json,.yaml. Override per project with.harness/file-size.conf.lint-on-edit.sh- Biome + ESLint on save for JS/TS; ruff check + format for Python.check-silent-errors.sh- blocks writes that introduce swallowed exceptions.block-dangerous-commands.sh- blocks force-push,git reset --hard, recursive rm on//~, before they run.rm-scope-guard.py- blocks anyrmwhose targets escape the working directory, and allows the rest. Complements the hook above, which only catches the catastrophic roots.check-codebase-health.sh- runs at session start. Reports files over 500 lines that need splitting. Silent when healthy.worktree-session-prompt.sh- at session start, says whether this is the shared main checkout or a linked worktree, and has Claude ask before editing the shared one. Two agents in one checkout is how a branch flip silently reverts a peer's edits.worktree-exit-offer.sh- on Stop, offers to leave the worktree once it is clean and fully pushed, which is the one moment exiting loses nothing.suggest-loop-improvements.sh- when you run/loop, proposes 2-3 tighter drop-in rewrites (explicit success criteria, stop condition, scope, verification) and lets you pick one via an interactive menu. Scoped to/looponly (client-side commands like/goalrun before Claude's turn, so they can't be gated).
Requires jq and python3.
This is the start of the file. Read the full SKILL.md on GitHub.
Security
- Security scanPassed, 100 of 100Pattern checks on the skill's files.
- Model reviewNot runA model review is scheduled for the next rescan.
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/new-project/SKILL.md
Popularity
- Repository stars
- 75
- Repository forks
- 4
- 7-day trend score
- 22
Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/sneg55/agent-starter/new-project)