Refine Prompt

Refine Prompt

Move a free-form prompt from "raw text" to "engine-ready" in one run. Produces reconstructed acceptance criteria, explicit assumptions, and a confidence band that decides whether the engine proceeds silently, halts for confirmation, or refuses to plan.

Sibling of refine-ticket — same reconstruction-of-intent pattern, different input shape. Tickets arrive structured (id, title, AC); prompts arrive as one string.

When to use

• The user invokes /work "<prompt>" or pastes a free-form request.
• The dispatcher hits input.kind="prompt" in the refine step.
• A prompt looks ambiguous, broad, or scope-undefined and the user asks "ist das klar genug, um loszulegen?".
• Before any plan/apply step on prompt-driven work — never after.

When NOT to use (near-misses)

refine-prompt is the first gate on prompt-driven flow. It does not plan, does not implement, does not write back anywhere.

Input

Exactly one path: a non-empty raw string carried in state.input.data.raw (built by work_engine/resolvers/prompt.ts). No branch detection, no URL parsing, no clipboard fallback — the calling command (/work) owns prompt capture; this skill only refines.

If raw is missing, empty, or whitespace-only the resolver already raised PromptResolverError. The skill never receives that input.

Modes and bypass

The skill honours prompt_optimization.inbound (or prompt_optimization.default when no inbound override is set) from .agent-project-settings.yml / .agent-settings.yml. Three modes:

Bypass prefix. If the raw prompt starts with the configured prompt_optimization.bypass_prefix (default /raw), the skill becomes a no-op regardless of mode. The dispatcher strips the prefix, passes the remainder through verbatim, and records bypass:true in the envelope so downstream surfaces (delivery report, --no-prose-synthesis) can attribute the skip.

/raw migrate auth.service.ts to use jose, keep the API shape

/raw is reserved at the prompt boundary only — it has no meaning mid-prompt and is not stripped when it appears inside the body.

Stack-config read (mini / max only)

When the mode is mini or max, step 3 may read these config files (read-only, scope-locked) to enrich the assumes: block:

• package.json — JS / TS framework detection (Next.js App vs Pages, Remix, SvelteKit, Astro, Expo, …)
• composer.json — PHP framework detection (Laravel, Symfony, framework-less)
• pyproject.toml / requirements.txt — Python framework detection
• CLAUDE.md / AGENTS.md — project-declared stack hints
• .cursorrules — project-declared stack hints
• tsconfig.json — TS path-alias / module-resolution hints

The skill MUST NOT read source files, .env*, secrets, or user data. Detection lands as a single assumes: stack=<framework>@<version> line; the medium-band halt is the user's chance to flip it.

Procedure

1. Read and analyze the prompt

Examine the raw text top to bottom before changing anything in state. Identify the single desired outcome in one sentence — verb + object + observable result. If the prompt names two unrelated outcomes (e.g. "fix login AND refactor the dashboard"), record both but flag scope-overload in step 5; the score will land in medium or low.

This is an analysis pass, not an execution pass. The skill does not modify the prompt, infer code changes, or call any tool — it investigates the input and produces a structured envelope the dispatcher reads.

2. Enumerate explicit constraints

Pull every concrete signal from the prompt verbatim:

• Files / modules named in the text (UserController, auth.service.ts, migrations/2024_…).
• Behaviour anchors — endpoints, routes, commands, fixtures.
• Hard rules — "must not break X", "without changing the API", "keep backwards compat".

Constraints come from the prompt only. Inferred constraints belong in step 3 (assumptions), never here.

3. Infer reasonable assumptions

Anything the prompt implies but does not state. Examples:

• "fix the login bug" → assumes the bug is in the existing auth/ module (no new auth provider).
• "add caching" → assumes the project's primary cache driver (per .agent-settings.yml / config/cache.php).
• "speed up the export" → assumes "faster" means runtime, not memory.

Each assumption is a single line, prefixed with assumes:. The medium-band halt surfaces them verbatim — no rewording, no explanations.

4. Generate the AC list

Three to seven bullet points. Each bullet is observable and testable in the project's existing test surface (Pest / Jest / pytest / etc.). Avoid:

• "works correctly" / "is fast" / "looks better" (untestable)
• "no regressions" (the test suite already covers that)
• "follows best practices" (not an AC)

Anchor each bullet to a constraint from step 2 or an assumption from step 3 — never both implicit.

5. Score confidence

Delegate to work_engine/scoring/confidence.ts:

import { score } from './work_engine/scoring/confidence.js';
const result = score({ raw: promptRaw, ac: reconstructedAc, assumptions });
// result.band ∈ {"high", "medium", "low"}
// result.score ∈ [0.0, 1.0]
// result.dimensions: Record<string, number>  // 0–2 per dimension
// result.reasons: string[]                   // human-readable rationale

The rubric (5 dimensions × 0–2, sum / 10) and band thresholds (high ≥ 0.8, medium 0.5–0.79, low < 0.5) are owned by confidence.ts. Do not re-derive them in prose.

6. Self-review (3-scan checklist)

Before emitting the envelope, run these three scans. Each is a fast pass; failure blocks emission.

1. Spec coverage — every concrete signal from step 2 (constraints) and step 3 (assumptions) is reflected somewhere in the AC list. Walk the constraint list top-to-bottom; each must anchor at least one AC bullet or appear in the Assumptions block.
2. Placeholder / TODO scan — the rendered envelope contains no <placeholder>, TODO, FIXME, tbd, ???, XXX strings. The literal angle-bracket placeholders in the template (<one sentence …>, <bullet>) must be replaced with concrete text before emission.
3. Type / shape consistency — every named file, module, route, or command in the AC matches the project's existing conventions. If the prompt names auth.service.ts but the codebase uses AuthService.php, surface the mismatch in Assumptions rather than adopting the prompt's spelling.

Source: adapted from an external reference.

Band-action mapping

The refine dispatcher step in directives/backend/refine.ts reads the score and returns one of three outcomes — the skill does not decide the action, only produces the inputs.

Output format

The skill emits a structured envelope; the dispatcher renders it. Required fields, in order:

1. Goal — single sentence, verb + object + observable result
2. Acceptance criteria — numbered list, 3–7 entries, each anchored to a step-2 constraint or a step-3 assumption
3. Assumptions — bullet list, each line prefixed assumes:
4. Confidence — band + score + per-dimension breakdown from work_engine.scoring.confidence

The shape below is the rendered surface for medium / low halts; for high the same envelope lands in the delivery report without a halt.

## Reconstructed prompt

**Goal:** <one sentence, verb + object + observable result>

**Acceptance criteria:**
1. <bullet>
2. <bullet>
3. <bullet>

**Assumptions:**
- assumes: <line>
- assumes: <line>

**Confidence:** medium (0.62) — goal_clarity 2 · scope_boundary 1 · ac_evidence 2 · stack_data 1 · reversibility 0

For low, the question replaces the AC list:

> The prompt does not name <missing dimension>.
>
> 1. <option that resolves the gap>
> 2. <alternative resolution>
> 3. <skip / abandon>

Gotchas

• The model invents AC that sound observable but aren't anchored in the prompt or a concrete file. Every AC must trace to a step-2 constraint or a step-3 assumption — no free-floating bullets.
• Assumptions are not commitments. The medium-band halt is the user's chance to flip them; the skill never asserts an assumption as fact.
• The scorer is heuristic, not LLM-based. Token count is not a signal — a 200-word prompt can score low if the goal is vague, and a 20-word prompt can score high if scope is unambiguous.
• UI-shaped prompts ("redesign the dashboard", "make the form prettier") score low on stack_data until R3 lands the UI directive set; band-action is a pointer to R3, not a refusal.

Do NOT

• Do NOT call this skill on input.kind="ticket" — that path runs through refine-ticket.
• Do NOT auto-confirm assumptions on the user's behalf in the medium-band halt. The halt is the contract.
• Do NOT stack multiple clarifying questions in the low-band halt. Iron Law: one question per turn.
• Do NOT mutate state.input.data.raw. The original prompt stays verbatim for replay; reconstructed output lands in data.reconstructed_ac and data.assumptions.
• Do NOT re-derive band thresholds in prose. They live in confidence.ts and only there.
• Do NOT read source files, .env*, secrets, or arbitrary user files when stack-detecting in mini / max mode. The allowlist above (package.json, composer.json, pyproject.toml, requirements.txt, CLAUDE.md, AGENTS.md, .cursorrules, tsconfig.json) is exhaustive.
• Do NOT strip the bypass_prefix mid-prompt. The prefix is only recognised at the prompt boundary; matches inside the body stay literal.
• Do NOT silently rewrite the prompt in max mode without surfacing the assumption list on a medium-band halt. The diff is the contract.

See also

• refine-ticket — sibling for ticket-shaped input
• prompt-optimizer — engine-outbound sibling; same prompt_optimization setting controls its mode
• work_engine/resolvers/prompt.ts — envelope builder
• work_engine/scoring/confidence.ts — rubric + band thresholds
• ask-when-uncertain — one-question-per-turn Iron Law
• artifact-drafting-protocol — this skill was drafted under it
• AI Council session: agents/runtime/council/responses/prompt-master-mini.json (2026-05-17) — analysis behind the mini/max split and /raw bypass