Questions about Recipaie
What Recipaie is
A JSON deployment recipe your AI assistant can follow and verify, step by step
Recipaie is a free, open specification for recipes an AI assistant can follow to install and configure software. Each step says what to do, why, when and how, plus how to verify it worked, so the assistant checks its own work instead of paraphrasing a README. A companion format, prodaie, records what happened; nothing here executes anything.
What people use it for
I need Redis on a new box. Can my AI do it from my notes?
Yes. Hand it the recipe at /example/redis and it follows six ordered steps, each with the command and its own verify check, so it confirms Redis answers before moving on. A variant that is already installed is covered by skip_if instead of being reinstalled. You end up with a running Redis and a prodaie record of what was done.
- Ask your AI
Deploy Redis here using the recipe on recipaie.com.
- What it does
- GET /example/redis → run 1-6 → verify → write prodaie
- What you get
- A verified install and its record.
I have install notes for my service. Can an AI follow them?
Write a recipaie: name, version, steps and step_count, where every step has order, what, why, when, how and verify. The why and when fields are what stop an assistant from improvising. POST it to /api/validate, or paste it into /validate, until it returns valid true; missing fields, a gap in order and a step_count that disagrees with the steps are all refused.
- Ask your AI
Turn these install notes into a validated recipe.
- What it does
- read /specification → draft JSON → POST /api/validate
- What you get
- A recipe your assistant can follow.
It keeps reinstalling packages already on the machine. Can I stop that?
Give the step a skip_if: a checkable condition such as a directory existing or a version already reporting itself. The assistant evaluates it before acting, and a skipped step is recorded in the prodaie as success true. verify still runs afterwards, so a skip is a claim the state is already correct rather than a shrug.
- Ask your AI
Don't reinstall what is already on this machine.
- What it does
- read skip_if → test it → skip or run → verify
- What you get
- Existing installs left untouched.
The same app is on three servers. What happened on each?
Record a prodaie: name plus locations, one per machine, each with path, status (complete, partial or failed), steps and step_count. Every step result adds success and details, so a failure says which step failed and what it printed instead of 'deployment failed'. The examples page pairs each recipe with such a record, including the same app deployed two different ways.
- Ask your AI
Record what happened on each of the three servers.
- What it does
- run per machine → add success and details → set status
- What you get
- One file showing each machine's result.
I want my AI to understand our stack before changing it.
A recipe works as context too: the why and when fields explain the reasoning behind each action, so an assistant can answer questions about the system or troubleshoot without running a command. The site states the format serves tools, processes or both, for execution, understanding or both. Point your assistant at the file and ask what a step does and why it exists.
- Ask your AI
Explain what step 3 does before you change anything.
- What it does
- read the JSON → answer from why and when
- What you get
- Answers grounded in your own steps.
Our runbook is a wiki page. Can it check itself?
Yes: turn each procedure into steps with a verify command, so the check is part of the document rather than a paragraph someone skips. Keep the human-facing page and put the recipe beside it; the schema is public, so a change either validates or does not. The field list is on /specification and four worked recipes are on /examples.
- Ask your AI
Convert this runbook into a recipe with a real check.
- What it does
- split into steps → add verify → validate
- What you get
- A runbook that checks itself.
A colleague sent a recipe. How do I check it first?
Validate it first: POST the document to /api/validate with schema_type recipaie, or paste it into /validate and read the errors. That covers required fields, sequential order and step_count, but not whether a command is safe, so still read the how of each step. The schema is public, so you can run the same check yourself with ajv.
- Ask your AI
Check this recipe before we run it.
- What it does
- POST /api/validate, read errors, read each how
- What you get
- A file known to be well-formed.
How to
How do I validate a recipe my AI just wrote?
POST it to https://recipaie.com/api/validate with {"schema_type":"recipaie","content":{…}} and read the answer: valid, name, version and the step count, or the errors. It is limited to 1 request per 5 seconds and 100 per hour per IP, and a 429 names the CLI fallback. The /validate page does the same in a browser.
How do I write a step my AI will actually verify?
Every step needs order, what, why, when, how and verify; only skip_if and comments are optional. Write verify as a command whose output decides the outcome, such as docker ps, systemctl is-active or a curl that must answer, rather than as 'check it works'. Order must run 1, 2, 3 without gaps and step_count must equal the number of steps, or the file is refused.
How do I record the outcome of a run?
Write a prodaie: name, plus at least one location with path, status (complete, partial or failed), steps and step_count. Each step result repeats the recipe fields and adds success and details; a deliberately skipped step is success true. Validate it with schema_type prodaie and keep the file beside the recipe so the next run has a baseline.
How do I check a recipe without sending it to your server?
Validate offline against the public schema: npx ajv validate -s https://recipaie.com/schemas/recipaie.schema.json -d your-file.json, or pip install jsonschema and use the same URL. No account and no rate limit. Fetch the schema once and keep it if the machine has no network.
How do I start from one of your examples?
Open /examples and pick one of the four recipes; each page shows the whole JSON, and most pair it with a prodaie record of a real run. tool-mcp is the shortest at five steps, so it is the clearest skeleton: copy it, replace the steps with yours and validate again. The /validate form's Load Example button inserts a two-step starter if you would rather begin in the browser.
How can my AI find its own way around this site?
Point it at /llms.txt: a curated map that names the specification, the validator, the examples and both schemas, and states in one line that the API takes a document and how it is rate-limited. From there it can fetch /specification for the field list and the schema URLs for the check. The example pages are HTML with the JSON on the page, so an assistant reads the page rather than a raw file.
Privacy and trust
Does your validator keep or change my document?
No. The endpoint parses the document, runs the schema check and answers in one response; there is no route that returns a submitted document, and the access log carries the path and status rather than the body. The limits are per IP: 1 request per 5 seconds and 100 per hour, answered with 429. If a document must not leave the machine, use the public schema with ajv or jsonschema instead.
Who makes this, and may I use the format in my own product?
Recipaie is an open specification published by Publifye AS (org.nr 826 774 622, Oslo), and the specification page states it is licensed under the MIT License. Both schemas are served as plain JSON at fixed paths, so you can read the format, validate against it and implement it yourself. Nothing on this site asks for an account.
Plans and trial
Is Recipaie free? Do I need an account, a key or a subscription?
Nothing to buy: this site has no store, no paid tier and no account system. Reading the specification, fetching the schemas and validating a document are all open, and the format is MIT-licensed. The only gate is the validator's rate limit, and the CLI check against the same schema has none.