Your first docs

1. It picks the doc set and the audiences. Pages go into the default doc set unless you pass --set=handbook. Without --audience, it asks who the docs are for.
2. It finds your agent. It looks for an agent CLI on your PATH and picks it. With several installed, it asks which one should write. --agent=codex chooses for you.
3. It prepares the agent. It makes sure the LaRecipe skills are installed in .claude/skills/ and .agents/skills/ (with Laravel Boost, Boost ships them instead). For a database set, it also adds the local larecipe MCP server to that agent's config, because the agent writes through it.
4. It shows the exact command and asks. You see the full command before anything runs.
5. The agent writes, you watch. For Claude Code, each step shows as a short line, such as > Read app/Http/Controllers/OrderController.php or > Write docs/api/orders.md, followed by what the agent says. Codex's own progress passes through as it is. There is no timeout; stop it at any time with Ctrl+C. If the agent stops with an error, LaRecipe reports its exit code. Nothing was committed.
6. It lists and validates the pages. LaRecipe compares the docs folder before and after the run (for a database set, the docs migrations folder) and lists each file as created or updated. Then it runs php artisan larecipe:validate on the set. When the pages pass, it reads "The new pages pass validation. Review them, then commit them with your code." When they do not, the command fails and lists the errors to fix first.

A repository set's pages are new files in your working tree. A database set's pages are applied to your local database and written to the branch's docs migration in database/migrations/docs/, so they reach production through the pull request like the rest of your change.

Interactively, larecipe:install asks "Generate starter docs from your app with your coding agent now?" as its last question, when it finds an agent installed. Or answer up front:

php artisan larecipe:install --onboard=api --onboard-agent=claude

Each audience is a preset in the larecipe-write-docs skill: an outline, a tone, what to read in your code and what never to write.

Product pages are task first and carry no class names or artisan commands. API pages are precise and scannable, with request and response examples taken from your Resources and tests. Package pages are copy-paste friendly, with configuration key names in prose. Internal pages are direct and dense, written for a Laravel developer on day one, with <Model> references for the data model.

A few spellings work too: users for product, api-consumers for api, package-users for package, and team or handbook for internal. Names it does not know are ignored.

For every audience, the skill tells the agent to leave out:

• Secrets and settings. No .env values, keys, tokens, passwords, connection strings, hostnames or internal URLs. A config key may appear by name, never its value. API examples use a fake placeholder such as YOUR_API_TOKEN.
• Customer data. No names, emails, IDs or records from seeders or fixtures that look real.
• Guesses. If the code does not say it, the page does not say it. The agent leaves a note in its report instead.

Per audience, it also skips internal routes and admin features on product pages, internal-only endpoints on API pages, and classes marked @internal on package pages.

The internal preset's handbook set should have access.mode set to auth or gate. Configuration basics shows how to add a second set.

The skill's onboarding workflow is the same for every agent. It reads AGENTS.md, README.md, composer.json and your existing docs first, and keeps the Introduction page the installer wrote, improving it rather than replacing it. It plans each page before writing, with the code each one will read. It groups the new pages by audience in the navigation, writes only inside the docs folder, runs the validator, and finishes with a list of the pages it wrote and anything it left out.

LaRecipe detects an agent by its CLI on your PATH (claude, codex, cursor-agent, gemini, copilot) or its files in the project (.claude, CLAUDE.md, .codex, .cursor, .gemini, GEMINI.md, .github/copilot-instructions.md, .github/skills). Onboarding picks only an agent whose CLI is installed, because it has to run it.

What you read is what runs: the command on screen and the process LaRecipe starts are built from the same arguments. For a repository set in docs/:

claude -p "Use the larecipe-write-docs skill to run onboarding for audiences: product, api (set: docs)" --allowedTools "Read,Grep,Glob,Edit(docs/**),Write(docs/**),Bash(php artisan larecipe:*)" --output-format stream-json --verbose

codex exec --sandbox workspace-write 'Use the $larecipe-write-docs skill to run onboarding for audiences: product, api (set: docs)'

Claude Code may read your whole project but write only inside docs/, and run only php artisan larecipe:* commands. For a database set, the allowed tools are Read,Grep,Glob,mcp__larecipe,Bash(php artisan larecipe:*): the agent writes only through the LaRecipe MCP author tools, which record each change in a docs migration.

Codex's sandbox cannot narrow writes to one folder, so LaRecipe says so under the command: review the diff before you commit. To write a database set, Codex also needs the larecipe MCP server in .codex/config.toml, in a project Codex trusts.

For Cursor, Gemini CLI and GitHub Copilot, open the agent in your project and send the prompt LaRecipe prints:

Use the larecipe-write-docs skill to run onboarding for audiences: product, api (set: docs)

LaRecipe never starts an agent without your yes:

• Interactively, it asks "Run it now?". Answer no and it prints the command to run later.
• --yes runs it without asking, for scripts.
• --print only prints the command.
• Without a terminal and without --yes, it prints the command and stops. That includes larecipe:install --no-interaction --onboard=api: the installer finishes, and you run php artisan larecipe:onboard --audience=api --yes when you are ready.
• larecipe:install --dry-run --onboard=api prints the command and runs nothing.

When you name an agent with --agent whose CLI is not on your PATH, LaRecipe prints the command and asks you to run it where the agent is installed.

When no agent CLI is installed, your docs keep the Introduction page the installer wrote, and LaRecipe prints the command to run once you have an agent:

php artisan larecipe:onboard --audience=api --set=docs

The workflow lives in the larecipe-write-docs skill, not in the command, so any agent that reads skills can run it from a chat: send it the prompt shown above. You can also write the first pages yourself; Markdown and components cover everything a page can hold.

Read live from this app's command registry:

php artisan larecipe:onboard [--audience [AUDIENCE]] [--set [SET]] [--agent [AGENT]] [--print] [--yes]