| name | spec |
| description | Create and maintain a specification that states what to build, not how. Use when writing a spec, editing requirements, or defining acceptance criteria. |
Spec
A spec states what must be true, never how to achieve it. Algorithms, specific API calls, data structures, byte offsets: those live in code or an architecture doc. The spec must be complete enough to reimplement from and precise enough to verify against. Every requirement is a claim someone can later prove or disprove.
Two modes, inferred from the request:
- New — draft a spec from a rough idea or a design conversation. Supply the structure so the user describes intent, not format.
- Edit — refine an existing spec: add requirements, strip the how that crept in, cross-check completeness against the code.
A spec need not arrive complete. It may start with a handful of FRs and grow one change at a time — the ID rules below make later growth safe.
When the source is an existing codebase rather than a design conversation, you are reverse-engineering the what from the how: the code says what the system does, never what the author intended. State each requirement as observed behavior and flag the ones you can't confirm as deliberate — a stopgap or a tuned default frozen as contract will drive future work toward preserving it. Mark the draft reverse-engineered and unverified against intent until a human confirms.
Labelled requirements
Every requirement gets a stable ID: a two-letter family prefix and a number (FR-12, NF-3, AC-9). IDs are the spec's addressing system. Acceptance criteria cite the requirements they verify, tests cite the requirement they cover, and cross-references point by ID, not by prose.
Core families. Add domain families as the project needs (ST styling, AU auth, SEC security):
- FR — functional: what the tool does.
- NF — non-functional: performance, reliability, error behavior, testing.
- AC — acceptance criteria: the conditions that together mean "done".
- D — deliverables: the artifacts that must ship.
- C — constraints: fixed decisions the builder may not revisit — language, runtime, distribution, hard exclusions (step 4 draws the line against mere conventions).
Rules:
- Never renumber. IDs are permanent addresses. Renumbering silently breaks every cross-reference and every test that cites one. Append new requirements with the next free number.
- Insert with a letter suffix (
FR-21a) only when a new item must sit beside a related one for reading order. Reordering for looks is not worth a broken reference.
- One claim per ID. A requirement with two independently-verifiable claims is two requirements.
- Retire in place. Removing a requirement never deletes its ID or renumbers around it. Mark it retired —
**FR-9** — *Retired (superseded by FR-38).* — so no citation ever dangles.
- Earn each family and each ID. A dimension gets its own family only when it carries distinct, independently-verifiable claims, not just functional surface that could sit under FR. Aim for a spec you could reimplement from, not an exhaustive inventory; when two readers would land far apart on the requirement count, you are enumerating mechanism, not contract.
Workflow
- Gather intent. For a new spec, read the reference the user points to and the conversation so far; for an edit, read the current spec in full. Resolve remaining ambiguity by asking one question at a time, each with your recommended answer, instead of guessing.
- Cross-check against code (when it exists). Every shipped behavior needs a requirement; every requirement needs a way to verify it. Enumerate the real surface from the source, not from memory: CLI flags, config format, error paths. A requirement the code contradicts is stale. Reconcile it.
- Draft from the template. Fill the section skeleton in
template.md, assigning IDs as you go. Done when every section is filled or explicitly marked not-applicable.
- Sort what from how. Any sentence naming an algorithm, a request shape, or an offset is how. Move it to the architecture doc and leave the requirement stating only the observable outcome. A fixed decision stays under Constraints only if reversing it would force a rewrite or change what the deliverable is — language, runtime, distribution form, hard exclusions like "no telemetry". A decision a tool could re-apply mechanically with every AC still passing — formatter, lint rules, indent width — is a convention: it lives in AGENTS.md, not the spec. A user-observable default (a default port, a timeout) is a requirement; the internal tuning value behind the same behavior is an open decision.
- Make edge cases requirements. The constructs a naive build gets wrong are requirements, not nice-to-haves. Give each its own ID so it earns its own test.
- Separate constraints from open decisions. Close with two explicit lists: what is fixed, and what is deliberately left to the builder. The silence between them is where scope disputes grow.
- Verify the traces close. Each AC maps to the FRs it exercises; each edge-case requirement appears in the testing section. At scale one AC cannot cite every requirement — ACs then cover representative bundles and the per-requirement trace is carried by test citations. A requirement reached by neither an AC nor a cited test, or an AC that verifies nothing, is the gap. Fix it before the spec is done.
Editing an existing spec
Before touching the living doc, think the change as a delta of IDs — what is added, modified, retired, and which traces it moves (which ACs gain or lose an FR):
ADDED FR-38; MODIFIED FR-12 (narrowed); AC-4 now also cites FR-38; RETIRED FR-9
Stable IDs and retire-in-place keep that delta legible in the git diff itself, so it needs no separate document — the spec stays a clean statement of the current contract. Load-bearing rationale goes in a PR description or an ADR, never as motivation prose in the spec. A tombstone's brief reason, and a policy's serves FR-x pointer, are the exception: permitted guards against silently re-adding what you removed, not motivation.
See also
plan — design and decompose once the spec is stablearchitecture-review, doc-review — where the how lives, and keeping the spec free of drifttemplate.md — the canonical section skeleton to copy from
Red flags
- Specifying how (an algorithm, a request shape, a data structure) where an outcome would do
- Drifting into PRD territory (user stories, success metrics, motivation) instead of testable requirements
- Renumbering existing IDs, or reusing a retired one
- Acceptance criteria that cite no requirement, or requirements no criterion verifies
- Edge cases written as aspirations instead of labelled requirements
- Fixed constraints and open decisions blended into the same prose
- A requirement stated so vaguely no test could prove it met
- Copying another project's domain families wholesale instead of choosing the ones this project needs
