init — Scaffold a coding-agent setup

Starting a new agent project means writing the instruction file, settings, hooks, and permissions by hand. Most people skip half of it. Init detects your stack and generates the full setup in one command.

claude-launchpad init
claude-launchpad init --harness cursor
claude-launchpad init --harness both

Default is Claude Code. --harness cursor writes native Cursor files (AGENTS.md, .cursor/). --harness both writes both surfaces and shared TASKS.md / BACKLOG.md once.

Flags

With --yes, /lp-enhance installation defaults to project scope (unless already installed). --yes without --force exits with code 1 when the selected instruction file already exists; combine -y --force for full automation.

What it writes (Claude Code)

Always writes these files, and creates additional files only when they do not already exist:

• CLAUDE.md: stack, commands, conventions, guardrails, backlog instructions, stop-and-swarm debugging rule
• TASKS.md: sprint tracking that stays concise (one-line completed sprints, max 3 session entries, ## Current Sprint empty between sprints). The SessionStart hook loads this every session so Claude picks up where you left off.
• BACKLOG.md: parked features using the 7-field WP template (Priority, Proposed, Stories/Docs, Depends on, Estimate, Trigger to pull, Definition of done). Structural P0/P1/P2/P3 sections plus a Changelog. See workflow for the full lifecycle.
• settings.json: schema, credential blocks (~/.ssh, ~/.aws, ~/.npmrc), bypass mode disabled, and hooks. Hooks run at 100% reliability (unlike CLAUDE.md rules at ~80%). The SessionStart hook loads TASKS.md on startup, resume, and after context compression (its matcher covers compact and clear). Includes: .env protection, destructive command blocking, auto-format, sprint review, SessionStart (new + resume), workflow-check, sprint-open-check.
• .gitignore: prevents local settings, plans, memory from being committed
• .claudeignore: language-specific ignore patterns
• lp-enhance skill: AI-powered CLAUDE.md improver. Installed during init in either project scope (.claude/skills/lp-enhance/SKILL.md) or global scope (~/.claude/skills/lp-enhance/SKILL.md) when not already present.
• rules/conventions.md: language-specific starter rules.
• rules/workflow.md: path-scoped BACKLOG/TASKS rules (YAML frontmatter paths: ["BACKLOG.md", "TASKS.md"]). Claude Code auto-loads it only when editing those files. See workflow.
• rules/hooks.md: path-scoped hook authoring rules (YAML frontmatter paths: [".claude/settings.json", ".claude/settings.local.json"]). Documents the input-via-stdin-not-env-vars contract, exit code 2 vs 1, multi-matcher caveats, isolation testing pattern. Auto-loads only when editing settings.
• hooks/sprint-size-check.sh and hooks/sprint-open-check.sh: warn on microsprints / oversized sprints / BACKLOG drift at commit time.
• hooks/workflow-check.sh: PostToolUse staleness check that warns on duplicate WP IDs across BACKLOG+TASKS, TASKS > 80 lines, Current Sprint > 15 items, Session Log > 3 entries.

What it writes (Cursor Agent)

init --harness cursor writes native Cursor files. Shared TASKS.md and BACKLOG.md are the same as Claude.

• AGENTS.md: stack, commands, conventions, guardrails — Cursor's instruction file.
• hooks.json: fail-closed security hooks (.env deny, destructive shell deny) plus workflow, format, and session-context scripts.
• .cursorignore: language-specific ignore patterns (Cursor's equivalent of .claudeignore).
• rules/*.mdc: path-scoped conventions, workflow, hook authoring, and verification rules (globs: frontmatter).
• lp-enhance skill: AI-powered AGENTS.md improver at .cursor/skills/lp-enhance/SKILL.md.

--force overwrites AGENTS.md only. Other Cursor files are created when missing and left alone if you already customized them.

Sprint review hook

Init generates a PostToolUse hook that triggers automated code review when a sprint completes. Without this, you commit whatever Claude produced. With it, Claude reviews its own changes before they land:

• Watches for edits to TASKS.md
• When all tasks are checked off (- [x]), outputs a review nudge
• Claude scans for dead code, debug logs, convention violations, and hardcoded values
• Works across long sessions with multiple sprints, even after context compaction

The nudge is a suggestion, not a gate. Claude can ignore it for docs-only or config-only changes.

Supported stacks

Languages: TypeScript, JavaScript, Python, Go, Ruby, Rust, Dart, PHP, Java, Kotlin, Swift, Elixir, C#

Frameworks: Next.js, FastAPI, Django, Rails, Laravel, Express, SvelteKit, Angular, NestJS, and 15+ more.

Next