microsoft/clarifying-scenarios

Purpose

Frontend rewrites in particular fail silently when the user prompt omits the target component library, screenshots, or compliance requirements. This skill catches those gaps before decomposition by:

1. Evaluating the raw user prompt + project facts against three kits (frontend, backend, generic).
2. Computing a deterministic clarity score.
3. Asking the user — exactly once per round, in a single batched form — for missing required/recommended fields.
4. Writing a canonical clarification.md consumed by Foundation/Design/Plan agents.

Inputs

You receive these from the coordinator (via task metadata or dependencyArtifacts):

| Input | Source | Required | |-------|--------|----------| | userInput (raw user prompt) | task metadata | yes | | classification (type, complexity, target, etc.) | coordinator §1 output | yes | | project_facts (detected tech stack, frontend/backend presence) | coordinator §2.1 output | yes | | interactive (boolean — is a TTY available?) | session env (default true) | no | | Existing clarification.md (if any) | {{BASE_PATH}}/clarification.md | no |

Outputs

Only one file is ever written to disk:

{{BASE_PATH}}/clarification.md   ← canonical artifact (written only on READY)

The clarification form is never persisted to disk. When the gate needs user input, the skill renders the form inline (as the body of its return value) and the coordinator embeds it in the [wait] message shown to the user. The user's reply comes back through the next invocation's userInput / conversation context — not by editing a file. This eliminates a class of bugs where a stale form lingers on disk after clarification completes.

Decision tokens returned to the coordinator:

• READY <path-to-clarification.md> — clarification.md written, ready to proceed
• NEEDS_INPUT — the body of this return value IS the form markdown; coordinator shows it to the user verbatim inside [wait]
• BLOCKED <reason> — inputs malformed or required info cannot be collected

Workflow

Step 1 — Scope detection

Apply the rules in references/scoring-rubric.md §"Step 1 — Scope Detection" to determine which kits apply:

• frontend — load references/kit-frontend.md
• backend — load references/kit-backend.md
• always — load references/kit-generic.md

If classification is direct or fix_bug, return READY with a minimal clarification.md that records scope: [] and clarity_score: 1.0. Do not run scoring.

Step 2 — Evidence extraction

For each field in each applicable kit, scan the userInput and project_facts for accepted evidence (per the kit's "Accepted evidence" list). Mark each field present or missing.

Be conservative: if you are unsure whether evidence is present, mark missing and let the user confirm. False positives are worse than asking.

Step 3 — Score & decide

Apply references/scoring-rubric.md §"Step 2-3" to compute per-kit scores and the overall pass/fail.

• All applicable kits pass → go to Step 6 (write clarification.md).
• Any kit fails → go to Step 4 (write form).

Step 4 — Generate structured questions

1. Load templates/clarification-form.md to identify the fields that need answers. Do not write anything to disk.
2. Remove fields for kits that are not in scope.
3. Pre-fill any fields that already have evidence in userInput or project_facts — mark them as prefilled: true with the detected value so the coordinator can show them as pre-selected defaults.
4. Return NEEDS_INPUT with the body as a JSON array of question objects. Each object has:

   {
     "id": "F2",
     "question": "What target component / UI library should the new frontend use?",
     "importance": "required",
     "options": ["shadcn/ui", "Material UI", "Ant Design", "Chakra UI"],
     "default": null,
     "prefilled": false,
     "prefilled_value": null
   }
   

   • id: field identifier (F1–F10, B1–B5, G1–G3).
   • question: human-readable question text (concise, one sentence).
   • importance: required | recommended | optional.
   • options: array of common choices (may be empty if free-text only). Include an "Other" option when applicable.
   • default: the default value if skipped (null for required fields).
   • prefilled: true if evidence was found in the user's input.
   • prefilled_value: the detected value (null if not prefilled).
5. The coordinator will present these questions interactively (not as a printed form) and re-invoke this skill with the collected answers.

Step 5 — Re-invocation after user reply

When re-invoked with the user's collected answers (formatted as "F1: <answer>, F2: <answer>, ..." or free-text):

1. Parse the user's reply against the field list. Tolerate variations: key-value pairs, a bullet list, or sentence form — all map to the same fields by id (F1–F10, B1–B5, G1–G3).
2. For each field:
   • If user provided an answer → resolution: user, value = answer.
   • If user selected a default / accepted pre-fill → resolution: user, value = confirmed value.
   • If answer is empty and field is not required → resolution: default, value = the field's default-if-skipped.
   • If answer is empty and field IS required → resolution: blocking, no value, append to blocking_gaps.
3. Re-score. If all required fields are now answered (or marked blocking), proceed to Step 6. Otherwise, if rounds < 2, return NEEDS_INPUT again with only the still-missing fields as structured questions.
4. If rounds == 2 and gaps remain, apply defaults to all non-required missing fields and mark unfilled required fields as blocking_gaps. Proceed to Step 6.

Step 6 — Write clarification.md

1. Copy templates/clarification.md to {{BASE_PATH}}/clarification.md.
2. Populate the YAML front-matter:
   • generated_at: current UTC time
   • scope: the detected scope list
   • clarity_score: the final overall score
   • rounds: number of clarification rounds taken
   • gaps: every field where resolution != user
   • blocking_gaps: every required field with resolution == blocking
3. Fill the body sections with the resolved values. Remove sections for out-of-scope kits.
4. Return READY <path>.

No cleanup step needed. Because the form was never written to disk, there is nothing to delete. This is the single source-of-truth file for clarification state.

Non-interactive / CLI mode

When interactive == false (no TTY, --yes flag, CI run):

• Skip Step 4/5 entirely.
• Apply default-if-skipped to every missing non-required field.
• For missing required fields: return BLOCKED missing required fields: <list>. The coordinator decides whether to fail or proceed with risks.

Round limit

Maximum 2 clarification rounds per session. After round 2, defaults are applied automatically — the user is not asked a third time.

Determinism guarantee

Given the same userInput, classification, and project_facts, this skill must produce the same clarity_score and the same set of asked fields. Do not introduce randomness or model-driven judgement at the scoring step — only at the evidence-extraction step (where light interpretation of natural language is unavoidable).

Resources

References

• references/kit-frontend.md — full frontend field set
• references/kit-backend.md — backend field set (v1 skeleton)
• references/kit-generic.md — baseline fields (always applied)
• references/scoring-rubric.md — scoring algorithm and pass thresholds

Templates

• templates/clarification-form.md — user-facing questionnaire
• templates/clarification.md — canonical artifact schema (clarification/v1)