Clarificar
Skill adriannoes/awesome-agentic-ai/cursor-claude-codex/skills/igoruehara-spec-driven/skills/clarificar
Use for a GRILLING SESSION — a relentless interview that turns diffuse intent into shared understanding before building. It walks the decision tree ONE question at a time, resolving dependencies between choices, always with a recommended answer; it explores the codebase/docs instead of asking when the answer already exists. Ideal at the spec gate (testable ACs / Definition of Ready) and the design gate (hard-to-reverse decision). It produces an understanding summary that feeds product.md/design.md/spec.md. Called by /nova-feature, /kickoff, /roadmap, and /validar, or trigger directly with /clarificar.From its SKILL.md
npx -y skills add adriannoes/awesome-agentic-ai --skill clarificarAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
SKILL.md
4.2 KB, 871 tokens by cl100k_base, as published. Nobody here has run it
Translation note: Originally authored in Portuguese (pt-BR) by Igor Uehara (igoruehara/spec-driven, MIT). Translated to English by this hub to keep the repository language consistent. Original content unchanged in meaning; see the upstream repo for the pt-BR source.
Skill: Clarify (plan/spec grilling)
A relentless interview to sharpen a plan, a spec, or a design before building. The target is
shared understanding: no ambiguity that turns into rework or SPEC_DEVIATION later.
Inspired by the grilling technique — adapted to the pipeline's vocabulary (testable ACs, tier,
hard-to-reverse decision → ADR, ubiquitous language).
When to use this and not "short batches": the rest of the pipeline asks in batches (
AskUserQuestion) for independent choices. The grilling is the opposite and the complement: for a decision tree with dependencies, where answer 3 only makes sense after answers 1 and 2. Use when the ambiguity is deep and branched, not when the options are orthogonal.
Principles (the engine)
- One question at a time. Ask the question, wait for the answer, and only then the next one. Several questions at once confuse and prevent each answer from refining the next.
- Walk the tree. Each answer opens/closes branches. Resolve dependencies in order: don't ask "how" before "whether", nor the detail before the boundary.
- Always propose a recommended answer. Don't interrogate in a vacuum — for each question give your
recommendation with the rationale (in
AskUserQuestion, the first option carries "(Recommended)"). - Explore before asking. If the answer is in the codebase, the docs (
specs/,docs/, ADRs, glossary) or a connected reference MCP, find it yourself — only ask what requires a human decision. (This is the knowledge verification ofCLAUDE.md.) Never invent: explicit uncertainty > confident guess. - Dig down to the testable. A vague answer ("fast", "secure", "several") does not close the branch — refine it until it becomes a verifiable criterion (a number, a concrete case, Given/When/Then).
Process
- Frame the target. State in 1-2 lines what is being grilled (this feature / this design / this priority) and what will count as "understood enough to proceed".
- Raise the open branches. From the existing material, mentally list the open decisions and the ambiguities. Prioritize the highest impact / hardest to reverse.
- Grill, one branch at a time. For each point: explore first; if a decision remains, ask one thing with the recommendation; integrate the answer; let it open the next branches. Repeat.
- Stop when the tree closes — no open branches that change what will be built. Don't extend for sport; once the user and you agree on the scope, end it.
Output — the understanding becomes an artifact
The grilling doesn't end in chat: consolidate the result to feed the pipeline.
- Understanding summary: closed decisions, non-goals that emerged, assumptions, and risks.
- Direct to the right destination according to what was clarified:
- became an acceptance criterion →
spec.md(Given/When/Then; multi-factor rule → Decision Matrix); - became a hard-to-reverse decision → becomes an ADR (
docs/architecture/adr/); - became a business term →
docs/glossary.md/domain.md; - left open on purpose →
docs/STATE.md(with the trigger to reconsider).
- became an acceptance criterion →
- Point to the next step of whoever called you (fill in the gate, resume
/nova-feature, etc.).
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.