Intent to spec

Input: an intent/NNNN-name/intent.md. Output: spec.md beside it, in the shape of intent/0000-template/spec.md. This is Stage 2, and its job is to apply the policies while the design is still free — the alternative is finding out in review, when the design is a diff and moving it is expensive.

Apply these while writing, not afterwards

Read the intent, then walk the constraints in this order, because they disqualify designs at very different costs:

1. Determinism (determinism-review). Does the proposed behaviour observe time, randomness or ordering? Everything does eventually — say how it reaches f_env::Env. A design that cannot is not a design yet.
2. The frame (frame-and-unsafe). Does this need unsafe? If so it lives in abi/, ring/ or kernel/, and if the intent puts it elsewhere the spec has to move it or argue for an RFC.
3. The licence boundary (licence-boundary). Does it touch imported code? Then the interface is a ring, and that shapes the whole design, not the last paragraph of it.
4. Evidence (claims-registry). What observation closes this? A named test, a claim with a threshold, or a document that can now say something it could not. Write it in the spec, because a spec whose evidence section is written afterwards gets the evidence that happened to be available.
5. Decisions (rfc-author). Does this contradict something written down, or would somebody reasonably re-litigate it? Name the RFC the change will need.

What a good spec does that a bad one does not

• Names what it is not doing, and where that work goes instead. This is the section that stops scope growing during the build.
• States the failure it expects. "Most likely to be wrong here" is a real section, and an empty one usually means nobody looked for it.
• Stays in the repository's terms. intent.md is written by whoever wanted the thing, possibly in their own vocabulary; the spec is where that becomes crate names, milestones and TODO.md IDs.
• Is shorter than the conversation that produced it, and longer than the intent.

Boundaries

The spec goes back to the originator before plan.md starts. Flag anything you had to decide on their behalf, in a list, at the end — those are the places a review is actually worth their time. Do not silently resolve an entry from the intent's Open questions: either answer it with a reason, or carry it forward.