You have a zip. Ten minutes from now you have an app
that boots and an agentic layer that works. Go in order. The env vars are the
only slow part, and /setup does most of it with you.
1. Unzip and install
bun installpackage.json pins packageManager to the one you picked, so everyone on the
repo installs with the same tool.
2. Run /setup in your agent
Open your coding agent in the folder: Claude Code, Codex, Cursor, Copilot or
any other. Run the setup skill: /setup in most agents, $setup in Codex.
In Gemini CLI, ask for the setup skill. It checks what is done, shows a plan,
and waits for your yes. Then it:
- creates
.env.local, generates the secrets and writes the defaults; - tells you which keys to paste, and where each one comes from;
- connects the MCP servers for your batteries;
- runs the database migrations once you paste the URL;
- sets your product name, tagline, domain and email, or runs a full interview.
It never asks for a key in chat. Keys go in .env.local, and it never reads
that file. Run it again any time: it starts from what is really there.
3. Or do it by hand
SETUP.md is the same plan, written for your exact picks. It lists every env
var your batteries need, where to get each one, and the order to set services
up. The order matters: a database before auth, auth before payments.
bun run setup:envThat creates .env.local, fills what a script can, and lists what is left.
You don't need every service live to boot.
4. Boot it
bun run devOpen http://localhost:3000. You get a landing page for a sample product, plus
/terms, /privacy and /llms.txt. All four read one file,
src/lib/site.ts. With a payments battery, /pricing renders too, with no keys
set.
The app should render. If it doesn't, that's a generator bug. Open an issue and
attach your agentic.config.json. It reproduces exactly what you generated.
5. Sign in to your own app
With an auth battery, run the database steps SETUP.md lists for your
picks (migrations first), then sign up at /sign-up. You land on /dashboard:
the signed-in app, with a sidebar, an account menu and settings for your profile
and security. Google, GitHub and Microsoft buttons appear once their keys are
set.
With the admin panel, make yourself an admin, then open /admin:
bun run admin:grant you@example.comThen give the landing page your real product. Ask your agent to run
/landing-copy with a short brief. It rewrites src/lib/site.ts and nothing
else.
6. Run verify
bun run verifyscripts/verify.ts is written for the services you picked. It checks each
required variable and pings each configured service, then prints what is
missing and where to get it. Green means the repo is wired up, not just
compiling.
7. Prove the guard hooks work
bun run verify:hooksPeople skip this step. Don't. scripts/verify-hooks.ts runs every guard,
battery guards included, through .agents/hooks/run.mjs, the launcher every
agent's hook config calls. It sends each guard real hook payloads, in Claude
Code's format and then in the format of each other agent the repo wires. Each
guard gets block cases and allow cases. Block cases run again from src/. A
guard that lets a block case through, or blocks an allow case, fails the
script.
Run it now, and again after anything touches .agents/hooks/ or an agent's
hook config. A guard you have never seen fire is a guard you are only assuming
is there.
The guards run in Claude Code, Codex, Cursor, GitHub Copilot, Antigravity,
Gemini CLI and Windsurf. Codex, Cursor and Gemini CLI run them only once you
trust the folder, and Codex asks you to review each one in /hooks. OpenCode,
JetBrains AI and Kiro run none. See the agentic layer.
8. Typecheck and test
bun run typecheck
bun run test
bun run test:e2eAll three should pass on a fresh repo. test:e2e runs the Playwright specs for
the pages your picks ship. A spec whose service isn't set up yet skips and says
what to set. Every pair in your pick is in the tested set,
and each preset runs whole in the matrix (see
how the generator works). So a failure here is worth
reporting.
9. Point your agent at the repo, then write the first plan
Open the repo in your coding agent. Ask it to read AGENTS.md and SETUP.md.
Then, instead of asking for a feature, run:
/ce-plan add team invites to the settings page
In Codex it is $ce-plan. The command comes from the Compound Engineering
plugin. Claude Code and Copilot offer it when you trust the folder. SETUP.md
has the install line for Codex, Cursor, Antigravity, Windsurf and OpenCode.
You get a plan in docs/plans/: the problem, the approach, the files, how to
verify it, the risks. Read it. Argue with it. That argument costs minutes now
and would cost a pull request later. When it's right, set status: approved
and approved: true in its frontmatter, then run /ce-work.
In team mode this order is enforced: plan-gate refuses
edits under src/ until an approved plan exists. In solo mode nothing stops
you skipping it. Try it once anyway, on something real, before you decide.
The ten-minute checklist
- Install finished
SETUP.mdread end to end.env.localcreated from.env.exampledevrenders the app- Signed up, signed in, and landed on
/dashboard verifyis greenverify:hookspasses: every hook blocks its block cases and allows its allow casestypecheck,testandtest:e2epass- One plan written, approved and worked
What to read next
docs/solutions/: the mistakes your batteries are known for, written before you make them. The same docs are public in the cookbook.AGENTS.md, and the nestedAGENTS.mdin each folder: what your agent has been told, scoped to the folders it applies to. Every rule in full is in.agents/rules/.- Designs and the token contract: before you change how anything looks.