How to use these recipes
Start with one useful job. Inspect the implementation, test the failure cases and understand what the evidence does not establish.
A recipe is a starting point, not a guarantee
These guides help a builder understand and implement a particular pattern. Some have evidence from an original product; others are drafts with small reference examples. Do not treat a badge as a security review or assume that an example has been tested in your stack.
A useful guide answers: what will this help someone do, what must already exist, which steps matter, how can the result be checked, and when should the pattern not be used?
Read the status first
| Status | What it means |
|---|---|
| Draft | Implementation, evidence or integration testing is incomplete. Read the stated limits. |
| In production | Evidence is reported from the original implementation. Independent reproduction is not established. |
| Field-tested | Another person has implemented the recipe and recorded the result. |
| Proven | Evidence exists from at least two independent implementations. It still does not guarantee your result. |
A chosen timeout or threshold is a configuration decision, not a measured outcome. A smoke test is not a production percentile. An offline helper test is not an MCP host test. Look for the date, versions, conditions, sample size and failure cases behind a claim.
The three layers
The guide explains the task, boundaries, decisions and tests.
SKILL.md gives a coding agent scoped instructions to adapt the pattern. It does not grant access, authorize a deploy or start an unattended process merely by being installed.
Reference code makes a specific part inspectable. A fixture or pure function may intentionally omit a database, authentication, scheduling or host UI. Those omissions should be explicit. Inspect the code before running it, use a synthetic environment first, and keep credentials out of prompts and logs.
Every guide shows its instructions and reference files in full. They all come as one download from naturate.io/recipes. There is no claim that an untested install command works.
A practical first run
Pick one recipe that matches an actual task. Try its smallest example. Break it deliberately with a missing input, denied permission, stale data or interrupted run. Then adapt only the relevant part to your service and repeat those checks there.
Keep consequential writes behind your service's real permissions and approval rules. Limit run time, spend and retries. Record the artifact produced, what passed, what failed and what is still unverified.
Public patterns, not private playbooks
The library shares individually useful patterns and independently written examples. It is not a dump of an organization's private prompts, operating rules or customer data. Everything committed to its public repository is public, even on a draft branch or if excluded from the website. Noindex is not access control.
The guides are licensed under CC BY-SA 4.0: share and adapt them with credit to Agent Recipes by Naturate, under the same licence. The skill files and reference code are licensed under Apache 2.0, so you can use them in your own product without sharing your changes. The Naturate name, logo and illustrations are not included.
Keep the implementation healthy
Every implemented recipe needs a way to detect regression: a contract test, source-freshness check, expected-output check or other measurable signal. Update the guide when that contract changes. Stop an automation that is producing activity rather than useful results.
Work with Naturate to adapt a pattern to a real product or workflow.