Proof
A Recipe has already done the explaining
The slow part of working with an agent is explaining what you want. A Vibedata Recipe carries the task, the acceptance contract and the guardrails, and a two-line pointer puts it into an Intent that already holds the context.
The Cookbook is live, and the first thing it changes is the size of the brief. Open a Recipe and the parts of a request that usually live in one engineer's head are already written down: the task, the acceptance conditions, the decisions the work has to ask for and the guardrails.
The walkthrough below runs the whole path. Find a Recipe, read Runs on and Verified by, copy the two-line pointer, and paste it into an Intent.
A Recipe is a requirement, not a chat prompt
A Recipe is one requirement-sized data-engineering outcome. It contains a task specification, an observable acceptance contract and guidance the agent reads when the Recipe is invoked.1
Three things on a Recipe page are easy to confuse. The Use this Recipe block is the invocation pointer: two lines that name the Recipe and its path and tell the agent to read it in the current Intent. The ## Prompt is the task specification. The ## Agent guidance says how the agent should approach execution.
The site shows the prompt and the guidance so you can evaluate the Recipe; neither is the copyable payload.1
That is why "just a prompt template" misses. A prompt is text you rewrite for this run. A Recipe is a versioned requirement with an acceptance contract, composed skills, the decisions it must ask for and explicit guardrails.12 The served identity of a Recipe is the Git commit a reader fetched; the catalog carries no revision of its own.1
The Cookbook narrows the search before you read a word
The Cookbook searches 146 Recipes: 139 Supported and 7 Planned at this reading. Search matches titles, triggers, descriptions, acceptance conditions and every label a Recipe carries. You can browse by job, by collection or by platform, and the readiness facet separates what runs today from what is planned.3
Search is the fast path when you know the job. The platform and collection views matter when you do not: a Recipe that cannot run on your target is a wasted read, and the platform filter removes it before you start.
Most Recipes are Supported. Proven is the higher bar: a dedicated Recipe-level eval proves the exact shape end to end, and the schema rejects Proven without one.1 Supported means the current capabilities can execute the Recipe end to end without that dedicated eval, so the two labels are not interchangeable.1
Readiness is a gate, not a badge
Planned means at least one required capability, integration or object type is unsupported today. The Recipe is published and discoverable, and it is not invocable: a consumer must show it as Planned and must not offer Use Recipe.1
That is why the facet goes before the title. A Planned Recipe can describe exactly the outcome you want and still be the wrong thing to paste today. It is a useful description of where the product is going; it is not a delegation you can make yet.
Runs on and Verified by are the two sections worth reading first
The worked example is api-to-bronze-incremental-contract, a Supported Recipe. Runs on names the execution targets: DuckDB, MotherDuck, Microsoft Fabric Lakehouse, Microsoft Fabric Warehouse, Redshift in Amazon SageMaker Unified Studio, and Databricks. Built with names dlt. Before you start carries the qualifiers that decide whether the Recipe fits the job.4
The qualifiers are specific. The Recipe was assessed end to end on DuckDB only; the other listed targets rely on dlt support and are unassessed. It needs a per-record cursor such as an update or creation time, and a resource without one is out of scope. Studio's rest_api connector is a stub, so the API needs another dlt route.4
Verified by is the section that tells you when the work is done. It lists the observable conditions that must be true before the Recipe is finished.4
- Every approved resource lands in bronze under its write disposition.
- Each load lands only the records the cursor selects.
- Refused records stay visible with a reason.
- Every record the source shows has one outcome, landed, refused or missed, and the outcomes reconcile.
- The result states what the load cannot capture.
Read those two sections before you copy anything. Runs on tells you whether the Recipe can execute where you work; Verified by tells you what you will be checking when it says it is done.
The pointer is two lines, and the Intent already holds the context
Here is the pointer for the worked example:
Use the Vibedata Recipe `api-to-bronze-incremental-contract` at https://github.com/accelerate-data/vibedata-cookbooks/blob/main/recipes/api-to-bronze-incremental-contract/recipe.md
Read the Recipe and execute it in the context of the current Intent.
Paste it into an Intent in Vibedata Studio. The Intent already holds the repository, the platform, the Domain, the sources and the approved requirements, so the Recipe does not need them restated. The Cookbook README puts the execution model in one line: Recipe definition + current Intent context = execution.1
The Recipe still asks for the semantic decisions Studio cannot supply. For the worked example that is the write disposition, the business key, the cursor and starting point, how deletes appear, what the schema contract does with a new or changed field, and where refused records go.4 You answer those in the Intent, once, and the Recipe reads them from there.
The reusable asset is the written requirement
A fair objection is that the work is bespoke and a catalog cannot cover it. A Recipe is requirement-sized, and a Collection groups related Recipes as a discovery view; it helps you find the right one, and it does not execute anything. The Recipe still asks the engineer to resolve the semantics it leaves open, while the Intent supplies the environment, the repository, the Domain and the sources.1
What gets reused is the written requirement. The model and the way you ask are inputs that change between runs. The Recipe is pre-written intent, and the same definition can execute in a different Intent without being rewritten.
The productivity frame this supports is the same team shipping more, measured by time to PR and time per PR. A Recipe targets the briefing before the first pull request; the review that follows is unchanged.
The limit is the decisions the Recipe cannot make for you
An acceptance contract can only check the conditions it names.1 If your job depends on a decision you have not made, the Recipe has nothing to check against, so it can finish without covering what you care about.
The next time a job lands, look for the Recipe before you write the brief. Pick one job you ran by hand this month, open the Cookbook, and read the Recipe that matches it. Would its Verified by list have saved you the briefing?
References
Footnotes
-
accelerate-data/vibedata-cookbooks,README.mdat2e17ae391c66a38248a5a15b1187be644e32b9dd: the Recipe definition, the execution model, the three prompt concepts, the readiness semantics and the platform values. The repository is the canonical definition of Recipes and Collections once they are materialized; the site and Studio consume its catalog. ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 -
accelerate-data/vibedata-cookbooks,docs/adr/0001-cookbook-recipe-contract.mdandschema/recipe.schema.jsonat2e17ae391c66a38248a5a15b1187be644e32b9dd: the accepted Recipe contract and the frontmatter schema, including the body rules underx-body. ↩ -
The Vibedata Cookbook,
https://www.getvibedata.ai/cookbook, read recently: 146 Recipes, 139 Supported and 7 Planned; search over titles, triggers, descriptions, acceptance conditions and every label a Recipe carries; browse by job, collection and platform, with a readiness facet. ↩ -
accelerate-data/vibedata-cookbooks,recipes/api-to-bronze-incremental-contract/recipe.mdat2e17ae391c66a38248a5a15b1187be644e32b9dd, and the live page athttps://www.getvibedata.ai/cookbook/api-to-bronze-incremental-contract, read recently: the Runs on targets, Built with dlt, Before you start qualifiers, Verified by conditions, Composes skills, Asks first decisions, Guardrails and the two-line Use this Recipe pointer. ↩ ↩2 ↩3 ↩4
Get the weekly roundup by email
One email a week: what got easier in Vibedata, a recipe to try yourself, and where data engineering is heading.
Or follow the RSS feed.
