| name | specification |
| description | Turn fuzzy intent into shared understanding and acceptance criteria — specification as a conversation, run one question at a time, before any story is split or planned. Use when an idea, feature request, or problem statement has no agreed rules or examples yet ("let's spec this out", "what should this actually do?", "we need acceptance criteria"). Produces an example map and acceptance criteria written back into the team's own story artifact, plus candidate glossary terms and parked questions. The agent-facilitated round is a first draft for a real conversation between humans — recommended, not optional. For decision trees with no artifact, see grill-me; for tightening an existing artifact, see find-gaps; for slicing agreed work, see story-splitting. |
Specification: the Conversation Is the Product
A specification is not a document. It is a conversation between stakeholders in which shared understanding gets built — and the written result is, at best, an educated guess about what is really required, refined until reality gets the deciding vote (Gorman). The value is in the planning, not the plan: asking an agent to "write the spec" while skipping the conversation misses the entire point. This skill exists to force the conversation — with you first, then between the humans who own the answers.
Where this sits: before story-splitting, before planning, before any acceptance test exists. Its output feeds all three.
| Resource | Load when... |
|---|
spec-review-template.html | Presenting a finished map for human review — the three-amigos handoff page |
references.md | Checking sources for the practices taught here |
The Contract
- Find the story's home first. Ask where this story lives — an issue, a file under
docs/stories/, a ticket, anywhere the team already keeps it. Everything this skill produces is written back into THAT artifact. If it has no home yet, ask where it should live; never invent a parallel convention silently.
- One question at a time, always with a recommendation. Never a questionnaire. Each question names its stakes, offers a recommended answer with reasoning, and accepts "park it" — parked questions get an owner and a date, never silence.
- Every assumption becomes a question. The moment you notice yourself deciding something the human never said, stop and ask it. An answer that changes nothing in the map was not a real question.
- The map is data, not prose. Build it as you go and keep it visible:
story: <one line, in the words the human used>
rules: # business policies / acceptance criteria
- rule: <policy in domain language>
examples:
- <concrete: real values, real outcome — never a restatement of the rule>
questions: # open assumptions attached to THIS rule (optional)
questions: # cross-cutting open assumptions, each with an owner once parked
acceptance criteria: # distilled from rules once examples stabilise
candidate terms: # vocabulary the conversation coined or contested
- Challenge the map before trusting it. For each rule: the counter-example question ("what would have to be true for this example to come out differently?"), zero–one–many, boundaries, and what-happens-when-it-fails. A rule with one example is a guess; a rule whose examples all agree is a hypothesis.
- Read the map's shape as a diagnostic. Question-dominated → not ready to build; keep talking, or park with owners and stop. Rule-dominated (more than ~6 rules) → too big; hand the map to
story-splitting. Balanced, with concrete examples per rule → ready.
The Recommended Path: Draft, Then Humans
The agent-facilitated conversation above produces a first draft of shared understanding — between the human at the keyboard and an agent. That is not yet a specification, because the people whose understanding must be shared were not in the room.
Recommended flow, multi-round:
- Round 1 (agent-facilitated): run the contract above with whoever brought the idea. Write the map back to the story artifact. Generate the review page (below) from it.
- Round 2 (humans, three-amigos style): the review page goes to the real conversation — business, development, testing perspectives together (Dinwiddie's three amigos). They talk; they annotate the page (every card takes comments); disagreements become new questions, not silent edits.
- Round 3+ (reconcile): their feedback returns to the agent round as answers and new red cards. Update the map and the artifact; regenerate the page. Repeat until a round produces no new rules, no changed examples, and no new questions.
- Then split: the stabilised map goes to
story-splitting; each child story carries its rules and examples onward — per-rule questions flatten into the single questions list that downstream mapping expects — and where an acceptance-test outer loop is installed, those examples become the seed of its per-slice mapping.
Reality still gets the final vote: acceptance criteria written here are hypotheses until the shipped slice confirms them. When reality disagrees, the conversation reopens — that is a feature of the method, not a failure of the spec.
The Review Page
For round 2, generate a single self-contained page from resources/spec-review-template.html: fill its JSON data slot (replace ALL occurrences of the double-underscore REVIEW_DATA token, escaping every < in string values as backslash-u003c), write to a temp directory, open locally, never commit the generated page. Structure the data as sections sharing one card engine: a Rules section (one card per rule — the rule as the card name, its concrete examples as the plain text, its attached questions in the highlight line, the verbatim map entry collapsible beneath), a Parked questions section (question as name, owner + context as plain text), and a Candidate vocabulary section (term as name, one-line gloss as plain text). Every card takes comments; the copy-feedback control assembles all annotations into one markdown block that returns to the next round. This page is a conversation artifact, not an approval record — regenerate it freely every round.
Vocabulary Capture
Specification conversations coin vocabulary constantly — that is half their value. Every term the conversation invents, contests, or uses in two different senses goes into candidate terms with a one-line gloss. Where the ubiquitous-language skill is installed, candidates enter its five-step language protocol individually (extraction gathers candidates; only the protocol admits them); where it is not, the candidate list still travels with the story so the naming conversation happens somewhere.
Boundaries
| Situation | Skill |
|---|
| A fuzzy decision tree, no artifact yet, resolving choices | grill-me |
| An existing story/plan/spec that needs holes poked | find-gaps |
| The agreed map is too big; slicing into child stories | story-splitting |
| Turning a child story into PR-sized plans | planning |
| Per-slice executable specification (where installed) | acceptance-testing |
| Naming the vocabulary the conversation surfaced (where installed) | ubiquitous-language |
Verification Checklist
