spec-author
This skill is the discipline of writing a spec that another agent can
turn into failing tests and another into passing code. It is invoked
from docs/graph/protocols/specify.md.
A spec is executable when every functional contract maps to at least one test in the suite and the test name names the contract. Specs that read like marketing or like implementations are not specs; implementation detail belongs in grill.md §8.
When to apply this skill
brainstormhas converged and the next step is to formalize the behavior.- An existing feature's contract is changing.
- A bug investigation revealed a contract that was implicit and is now being made explicit.
- An ADR introduces a new system behavior that needs a spec.
How to write each section
§1 Summary
One paragraph: what this spec covers and why, in the two sentences you would give the next agent on the team before they implement it.
§2 Scope
Two bullet lists: "in scope" and "out of scope". The out-of-scope bullets are equally important. Reading the spec a year later, the out-of-scope list is what tells the next agent "no, we considered that and excluded it deliberately."
§3 User-facing behavior (product)
Describe what the user experiences. Use the user's vocabulary, not the system's. If the spec is for an internal API or a job, the "user" might be another service or a developer: name them and describe their experience.
§4 Functional contracts (architect)
The heart of the spec. Each contract is one Given/When/Then, single- outcome, observable from outside.
Naming. Use UPPER_SNAKE_CASE slugs that read as sentences. Tester
turns these into test names.
### Contract: SUBMIT_VALID_FORM_RETURNS_2XX ### Contract: SUBMIT_FORM_SCHEMA_INVALID ### Contract: SUBMIT_FORM_PERSISTS_RECORD ### Contract: GET_SUBMITTED_RECORD_RETURNS_BY_ID
The ### heading form is load-bearing: spec-lint.py only counts
### Contract: SLUG headings as live contracts, so a bare Contract:
line is invisible to coverage.
One outcome per contract. "Returns 201 and sends an email" is two contracts: one for the response, one for the side effect. Many small, single-outcome contracts compose into the behavior.
Observable from outside. If the test has to inspect a private field, the contract is wrong. Move the observation to a public surface (a returned value, a queried record, an emitted event).
§5 Non-functional requirements
Only the constraints that bind this behavior; the template carries the category list and the bind-only rule at point of use.
§6 Data shapes (architect)
Schemas for inputs, outputs, persisted state. Use a language- agnostic notation (YAML-like) by default; cross-link to native schema files when they exist.
Required fields, types, allowed values, max sizes. These become test fixtures and validation rules.
§7 Failure modes (architect, with security)
For each contract, the named ways it can fail and what happens. This is what separates a spec from a description: a happy-path-only spec is half a spec.
Failure: SUBMIT_FORM_SCHEMA_INVALID - Trigger: payload violates §6 schema - Response: 422 with field-level error in body - Side effects: audit log entry; nothing persisted - Recovery: client may resubmit with corrections
Security adds adversarial cases here: prompt injection, tool hijacking, data exfiltration when AI is involved.
§8 Examples
Concrete input/output pairs. Three minimum: one happy, one edge, one
failure. Examples are the seed for test fixtures and turn abstract
contracts into concrete tests, so make them real values, not
<placeholders>.
§9 Acceptance criteria (product)
Measurable conditions for "done". Each criterion maps to one or more contracts in §4 and to one or more tests in §10.
Acceptance criteria include the non-functional ones (latency, accessibility). Write them now, beside the functional ones, so the tester can encode them.
§10 Test mapping (tester)
A table mapping each contract and each acceptance criterion to the tests
that cover them. Update as tests are written. Status values: pending
(no test yet), red (test exists, fails), green (test exists,
passes), skipped (with reason).
§11 Open questions
Every "we'll decide later" with a named resolution path. A spec with
open questions stays in draft.
The sign-off rule
A spec is signed when every sign-off below that applies is recorded; until then it is unsigned and nobody plans against it:
- product ✓ confirms §3 and §9 reflect the user outcome.
- architect ✓ confirms §4, §6, §7 are coherent.
- tester ✓ confirms every contract in §4 is testable (the testability review).
- security ✓ if the spec touches auth, secrets, payments, uploads, external integrations, or AI behaviors that act on data.
Sign-off goes in §0 (Metadata), and the signed spec stays draft.
Sign-off says the contract is authoritative; active says it is under
test. The signed spec moves to active in the change that lands its
first RED tests (test-first's COMMIT; verify.status-evidence owns
the moment), and to implemented when every contract is green. A live
status over an empty assertion set is a false green: active with no
test is a red gate until RED lands, and a false green once someone
silences it.
A spec that grows by slice (specify.flow) is signed the same way, one
slice at a time. The slice's contracts, their §10 rows (status
pending) and a §12 changelog row naming its signers are written before the tester writes the slice's RED, and
they land in the same commit as that RED, so a live spec never holds a
contract with no test.
Before a spec is signed, its prose passes
docs/graph/skills/humanizer.md in file mode and
python3 docs/graph/prose-lint.py --file <spec> --against HEAD reports
no strong tell and no dropped contract slug, number, or code span.
The sign-offs are judgment; the shape and the coverage are mechanical.
python3 docs/graph/spec-lint.py (the kernel §3.1 gate in the gate
table of docs/graph/protocols/verify.md) checks every spec's shape
(unique slugs, a §10 row per contract, §9 criteria that map to real
slugs, sign-offs present on anything past draft) and, for live specs,
that every slug appears in at least one test. A draft is shape-checked
and not counted for coverage, so a spec in authoring never reports
uncovered.
Spec drift management
When the code and the spec disagree, decide deliberately which side is right (kernel §4 owns the boundary):
- Code is right, spec is wrong: edit the spec deliberately, add a §12 changelog entry, get re-sign-off.
- Spec is right, code is wrong: file a bug, add a regression test, fix the code.
- Both partially right: the brainstorm needs to revisit the
contract; back up to
brainstormorspecifyand write a new version.
Reference files
docs/graph/templates/spec.template.md: the template.docs/graph/protocols/specify.md: the protocol.docs/graph/protocols/test-first.md: what happens next.docs/graph/agents/01-architect.md,docs/graph/agents/04-tester.md,docs/graph/agents/08-product.md: the three co-authors.
