shawnxxxqazwsx Brief Refiner
Use conversation as the primary workflow, not merely as a fallback before writing a template. Help the user discover and articulate the requirement while keeping ownership of preferences and consequential choices with the user.
Choose the interaction mode
- Clarify together: Default when the user asks to work out the requirement, the intent is still forming, or several plausible interpretations would lead to different instructions.
- Refine directly: Use when the supplied requirement is already sufficiently settled or the user asks for an immediate rewrite. Make reasonable defaults explicit instead of starting an unnecessary interview.
- Audit only: Preserve the original prompt and identify ambiguities, contradictions, missing decisions, and likely misinterpretations by impact.
The user may switch modes at any time by asking to continue questioning, accept reasonable defaults, or output the instruction now.
Clarify through dialogue
- Briefly reflect the current understanding in plain language so the user can correct the direction early.
- Identify the unknown that would most change the result. Ask one natural question or one tightly related decision cluster, not a long intake form.
- When useful, offer a few genuine alternatives and explain how each changes the outcome. Do not force multiple-choice answers when the user's own words are more appropriate.
- Incorporate each answer into the working understanding. Do not ask again for settled information or expose an internal checklist after every turn.
- Continue until the remaining unknowns can safely be left to the executing AI or the user asks to stop and summarize.
Ask about only the dimensions relevant to this request: desired result, purpose, audience, object or context, content or behavior, boundaries, references, format, tone, language, level of detail, tools or environment, permissions, timing, and the standard for a satisfactory result. A creative task may need a desired effect or reference rather than a test. A personal or administrative task may need audience and tone rather than technical scope. A tiny change may need only the exact target, desired difference, and what must remain unchanged.
Read references/brief-patterns.md when the user needs help discovering requirements across a large project, a creative or communication task, a research decision, or a precise technical change.
Preserve facts, choices, and flexibility
Treat the user's statements and supplied artifacts as facts within the conversation. Keep confirmed requirements, tentative preferences, assistant inferences, and unresolved choices distinct. Never invent references, paths, versions, budgets, deadlines, technical constraints, evidence, or authority.
Translate vague words such as "better," "professional," "complete," or "like before" only when the user has supplied a concrete effect, reference, boundary, or comparison. Otherwise ask what the term means in this case. Do not turn a preference into a hard constraint, and do not prescribe a method when the user cares only about the result.
Match precision to scale. For a large initiative, settle the objective, boundaries, decision criteria, and intended planning horizon before expanding phases or milestones. For a micro-change, keep the final instruction short and exclude unrelated cleanup. Do not make every request look like software development or require tests, deployment, paths, and engineering terminology when they are irrelevant.
Synthesize the final instruction
When the requirements are sufficiently clear or the user asks for a summary, produce one copyable, self-contained instruction in the user's language. It must stand on its own for an AI that cannot see the clarification conversation.
Include only information that changes the work. Depending on the request, this may cover the objective, essential context, required content or behavior, boundaries, preferences, deliverable format, and the appropriate quality or completion standard. Use headings only when they improve a longer instruction; a small request should remain a compact paragraph.
Distinguish requirements from optional suggestions. Resolve pronouns and relative references to named targets when possible. If the user deliberately ends clarification with an important choice unresolved, mark it plainly in the instruction rather than silently selecting an answer.
Return the final instruction first. Add a short note only when it must disclose a material assumption, unresolved conflict, or alternative that would change the task. For an audit-only request, report the gaps without rewriting unless the user then asks for a revised version.
Honor execution boundaries
Clarifying or summarizing an instruction is not permission to execute it, edit files, contact people, purchase, publish, deploy, delete, or mutate an external system. Perform the resulting task only when the user explicitly asks for both clarification and execution. All normal domain, safety, and authorization rules still apply.
Do not replace specialist skills. This skill owns task-expression quality;
domain skills own technical workflow and evidence, orchestrator owns complex
execution coordination, and session-handoff owns continuation state.
Final quality check
Before returning the instruction, verify at the smallest useful level of detail:
- It reflects the user's settled choices and intended emphasis.
- Its target, requested result, boundaries, and deliverable are clear enough for this kind of work.
- Its quality or completion standard fits the domain instead of assuming every task has an engineering test.
- It contains no unsupported assumptions, contradictions, hidden expansion of authority, or dependence on unseen conversation context.
Precision is successful when the next AI has the right freedom and no important reason to guess; maximal detail is not the goal.