A little guidance. Then get building.

What is a recipe?

A recipe is a small piece of reusable engineering guidance. It explains when to use a pattern, how to implement it, what can go wrong, and how to check your work.

It might cover one detail—like writing a file safely—or combine several patterns into a component. You choose how much guidance to bring into your project.

Find a recipe

To use Sous from your coding agent, install the prebuilt client with one command:

curl -fsSL https://sous.soundheart.io/install.sh | sh

1. Open Recipes, or use the client in your coding agent.
2. Search with a few technical terms, such as retry jitter or UTF-8 byte budget.
3. Read the recipe’s use cases and limitations before choosing it.

If a search comes up empty, try shorter terms or another name for the pattern. Avoid putting private app names or customer details into a query.

$sv search utf8 byte budget

Codex example. In Claude Code and Pi, use /sv.

Use it with your agent

Ask your agent to inspect the recipe and apply the relevant guidance to your task. Check the language, library versions, and assumptions together.

A recipe guides the implementation; your agent still works with your code. Run the suggested checks in your own environment.

Need several patterns together? The Workbench lets you compose recipes and prepare a context bundle. Start with one recipe when that is enough.

Share a useful lesson

Once a solution is worth reusing, ask your agent to prepare a recipe. You’ll need a Sous contribution token; the client setup guide explains where to put it.

1. Generalize it locally. Keep the technical mechanism, public libraries, version assumptions, and useful checks.
2. Remove identifying details. No app names, private paths, internal URLs, customer information, secrets, or descriptions that reveal your product.
3. Review before publishing. Read the exact proposed recipe and approve it when you’re happy to share.

The client starts in manual mode. Hooks are optional and require explicit opt-in; they should never upload your coding transcript or source files.

Found a problem in someone’s recipe? Field notes are for failures, compatibility details, and evidence that can improve the next version.

Know what to trust

Sources and maturity help you evaluate guidance. Generated means the recipe is not independently verified. Proposed tests are things to run, not proof that someone ran them.

Check whether the documented versions match your stack. A published recipe version stays fixed, so you can point back to the exact guidance you used. That doesn’t guarantee an AI will generate identical code.

Common questions