A Claude Code skill · /repo-bootstrap

Asks first. Proposes with evidence. Generates only after your go.

Analyzes any repo and sets up a production-grade AI agent workflow: agents, skills, hooks, architecture tests, lint guardrails, CI, permissions, MCPs and rulesync. For new and existing projects.

Copy the prompt for your agent Read the README

Tell your agentClaude Code · Codex · Cursor · OpenCode

Install the agentic-repo-kit skill by following https://github.com/rwxce/agentic-repo-kit/blob/main/INSTALL.md

Your agent asks one question, copies the skill and tells you when /repo-bootstrap is ready.

Install manually

git clone https://github.com/rwxce/agentic-repo-kit.git
cp -R agentic-repo-kit/skills/repo-bootstrap ~/.claude/skills/

For one project, copy the folder into <your-repo>/.claude/skills/ instead. Other tools that read SKILL.md-style skills can use the same folder.

01 / The problem

Coding agents start coding before they plan.

They build custom engineering when an SDK or import exists, miss cross-system side effects, call work done after mocked tests pass, and pile up bloated config. Every guardrail in this kit was added after a concrete failure in a real SaaS codebase.

3–6 of 18

side effects missed by each planner without the protocol. None checked the provider docs.

~17 of 18

side effects found with the full pipeline.

A written protocol beats a bigger model.

A cheaper model with the planning protocol found nearly as many side effects as the most expensive one.

02 / How it works

Nothing is written until you say go.

Phase 0

Scan and ask

Scans the repo and shows an options menu: goal, families, tools, code style, depth and output.

Mode

New or existing

A clean project starts from the templates. An existing one gets a closer look first.

Existing repos

Discover and benchmark

Discovery lenses, a code-style decision, and a big-feature benchmark run in fresh contexts.

Phase 4

Propose with evidence

A proposal with file:line evidence for every artifact. Then it waits for your go.

Phase 5

Generate PRs

Parallel subagents turn the approved proposal into pull requests.

Phase 6

Verify for real

Tests, real hook triggers, CI runs, benchmark re-runs and adversarial review.

/repo-bootstrap Asks what you want, then proposes.

/repo-bootstrap quick Skips the benchmark.

/repo-bootstrap audit Audits and improves an existing agent config.

03 / What it can set up

Eleven families. Pick all or one.

Works with Claude Code, Cursor, Codex CLI, OpenCode, Copilot, Gemini CLI and others, with rulesync as a single source or native files.

Root rules

Plan before code, docs first, pinned dependencies, full-circle verification.

Agents

Orchestrator, implementer and reviewer roles, per-domain implementers.

Skills

plan-feature, verify-integration, reflect, debugging, pr-review.

Hooks

Outward-writes blocker, dangerous-command gate, lint on change, PR preflight.

Architecture tests

Registry and filesystem sync, unique codes, i18n parity.

Custom lint

Layer boundaries, no cycles, SDK gateways, colocated tests.

CI

AI-config sync, lint, typecheck and tests, parity checks, branch-policy guard.

Automations

Workspace setup, release promotion, vendored-skill version checks.

Permissions

Deny, ask, allow layers for shell, secrets and outward actions.

Registries

Side-effect registry, blueprint inventory, environment variables.

MCP servers

Official servers with a prod and sandbox split. Proposed, never installed without approval.

04 / Questions

Before you run it.

Does it work without rulesync?

Yes. Choose “native files only” in Phase 0, and it writes .claude/ (or your tool’s) files directly.

Will it rewrite my code?

Only if you choose the behavior-preserving refactor. Even then it produces a separate plan, with characterization tests first and small PRs.

How much does it cost?

A quick run is roughly 60–120k tokens. The full benchmark adds roughly 200–300k. The benchmark is the step that finds real gaps, so it is recommended for repos that handle money, auth or personal data.

Does it install MCP servers?

No. It proposes official servers with their install commands and waits for your OK on each one.