AI Developer Cookbook

The AI Developer Cookbook is a set of copy-paste prompts for Claude Code, Codex, and Cursor, grouped by stack: frontend, backend, databases, data engineering, DevOps, mobile, and error handling. Each recipe pairs a prompt with the machine check that proves the result: a failing-first test, a type check, or a build, so you verify behaviour instead of reading every line.

This hub is for developers and tech leads who already have an agent installed (start here as a developer) and need the exact words for a real task. You have seen the pattern: the agent says “done”, the diff is 600 lines, and you cannot tell whether it works without reading all of it. The recipes fix that by asking the agent for its evidence up front.

Which cookbook section fits your task?

Pick the section for your stack. Each section hub has a table that maps frameworks to recipe pages.

Your task	Section	What its recipes prove the output with
React, Next.js, Vue, Nuxt, Angular UI	Frontend recipes	Type, test, accessibility, and screenshot gates
APIs and services in Node.js, Python, Go, Java, .NET, Rails, PHP, Rust, Elixir, serverless	Backend recipes	Integration tests on a real database, race detectors, contract diffs
Migrations, ORMs, SQL tuning, NoSQL	Database recipes	The change run on a real engine, before and after query plans
dbt models, SQL pipelines, notebooks	Data engineering recipes	A dbt build with data tests on a dev target
Docker, Kubernetes, CI/CD, IaC, monitoring	DevOps recipes	Linters, dry runs, and plan output before anything applies
Expo, React Native, Flutter, native iOS and Android	Mobile development	Simulator and emulator runs, UI-test evidence
Debugging, logging, recovery	Error handling recipes	A reproduction test that fails before the fix and passes after it

Every recipe follows the same shape: the scenario, the prompt in a tip box, the check that proves it, and what breaks. The recipes are self-contained, so read them in any order.

Prepare your tool before the first recipe

The recipes assume the agent knows your build, test, and lint commands. Without them, “run the tests” becomes a guess. Give the agent an instruction file first:

Claude Code

Run claude in the repository root, then /init to draft a CLAUDE.md. Paste recipes at the prompt. For large changes, start in plan mode (/plan) and approve the plan before any file changes.

Codex

Run codex in the repository root, then /init to draft an AGENTS.md. Codex reads AGENTS.md before it starts work, but since Codex 0.150.0 an untrusted project does not supply it, so trust the repository when Codex asks. /plan works the same way as in Claude Code.

Cursor

Open the agent panel and paste the recipe there. Put the commands in a project rule; Cursor project rules covers the format and how to make a rule apply to every request. Use Plan Mode for multi-file changes.

Start on your tool’s default model. As of 2026-10-12 that is Claude Opus 5.5 in Claude Code from v2.1.280 (both release channels) and GPT-6.1 Sol in Codex (bundled default since CLI 0.159.1). Raise the effort level first (/effort in Claude Code; /model in Codex sets both model and reasoning effort) before you switch models, and see the models hub for when a switch pays off.

Copy-paste prompt: give the agent your gates before any recipe

Read the build files, lockfile, CI workflows and test config of this repository.
Write a "Commands" section for the agent instruction file with the exact
type-check, lint, test, single-test and build commands this repo uses today,
and a "Definition of done" section: all of them pass and your reply ends with
the last lines of each. Then run every command you listed and report which
ones failed or need a service (Docker, a database) that is not running. Do not
change any source file in this step.

How do you verify a recipe’s output without reading every line?

Ask for proof, not for code. A recipe’s result is trustworthy when three things hold:

1. A test failed before the change and passes after it. A test written after the code often asserts whatever the code happens to do.
2. The gates from your instruction file pass, and the agent pastes their output.
3. CI runs the same commands on the pull request as a required check. The instruction file is advice; CI decides. See CI/CD pipelines.

You still sign off on two things: the acceptance criteria before the agent starts, and irreversible steps after it finishes, such as a migration that drops a column. Collect both in the pull request’s evidence bundle.

Copy-paste prompt: wrap any recipe in a proof requirement

Before you change any code for the task above, write the tests that define
"done" and run them to show they fail. Then implement until they pass. Do not
edit or delete an existing test to make it pass; ask me first. Finish with a
summary that lists: each new test and what it proves, the last lines of every
command in the Definition of done, and anything you could not verify.

What goes wrong when you paste a recipe?

• The prompt names a stack you do not use. The agent invents files to match it. Recovery: replace the framework, paths, and commands with yours before you paste, and keep the proof steps unchanged.
• The agent mocks the thing under test. The tests pass against a fake database or a stubbed browser. Recovery: require a real engine in the Definition of done, as in integration testing.
• The agent edits a test until it goes green. Recovery: forbid it in the instruction file, review the test diff separately, and read protect the oracle.
• The code uses an API from an older framework version. Recovery: give the agent current documentation through Context7 or the framework’s own MCP server where the recipe names one.
• The agent reports “done” without running the gates. Recovery: make the CI job a required check, and in Claude Code add a Stop hook that runs the tests (hooks automation).

Where to go next with the cookbook

Backend recipes

Frontend recipes

Writing acceptance criteria an agent can check

Making a codebase agent-ready

Start here as a developer