--- name: interactively-refine-idea description: >- Turn a vague research idea seed into a clearer research direction through clarification, axes, assumptions, risks, and next-step recommendations. Use when a user has an early idea and needs an idea partner rather than a list of ten brainstormed topics. Output is interactive guidance with questions and 2-3 concrete directions. metadata: owner_model: ParadoxGPT-Creator-4B skill_family: creator version: "0.1" output_type: interactive default_language: zh trigger_keywords: [idea seed, refine idea, research direction, clarify, brainstorm, 交互, 科研想法, 创新] required_inputs: [user_idea_seed] optional_inputs: [constraints, research_context, target_venue, user_goal] handoff_to: [generate-idea-variants, identify-fatal-concerns, map-claims-to-experiments, build-intro-argument-chain] --- # Interactively Refine Idea ## Purpose 和用户一起把模糊想法推进成可写论文的 research direction。Creator 不应立刻生成十个点子,而要先澄清研究空间、关键假设、约束条件、风险和可验证路径。 ## When to Use - 用户说“我想做 X,但不知道怎么创新”。 - 用户有一个 seed,但没有清楚 gap、insight 或实验路径。 - 需要在 novelty、feasibility、evaluation risk 之间做取舍。 ## Do Not Use - 输入是一篇完整论文,要反推 idea → 用 `reverse-engineer-paper-idea`。 - 用户已经有明确 seed,只需要多个版本 → 用 `generate-idea-variants`。 - 需要完整实验计划 → 用 `map-claims-to-experiments` 或 Designer skill。 ## Required Inputs - `user_idea_seed`: 用户的模糊想法。 - `constraints`: optional,预算、模型规模、数据、时间、目标会议。 - `research_context`: optional,相关工作或已知背景。 - `target_venue`: optional。 ## Output Contract 输出必须至少包含: 1. **clarifying_questions** - 先问 3-5 个关键问题。 2. **research_axes** - 把方向拆成可选择的轴。 3. **current_risk_reading** - novelty、feasibility、evaluation、claim risk。 4. **candidate_directions** - 2-3 个可推进方向,每个含 insight、method sketch、test。 5. **recommended_next_step** - 下一步最值得做什么。 ## Procedure 1. 先复述 seed 的核心意图,不急着生成 idea。 2. 拆研究轴:对象、机制、评估、数据、约束、失败模式。 3. 找最可能被 reviewer 批的点,尤其是 novelty 和 evaluation。 4. 生成 2-3 个方向,每个方向必须有 tension、core insight 和验证路径。 5. 推荐一个下一步,说明为什么不是另外几个。 ## Quality Bar 一个好的输出必须: - 有交互感:先澄清,再提出方向。 - 不列 10 个泛泛 topic。 - 每个方向都能解释为什么 nontrivial 和怎么测。 - 推荐下一步考虑用户 constraints。 一个差的输出: - 直接堆 brainstorm 列表。 - 只追求 novelty,忽略可实验性。 - 对 reviewer risk 没有预判。 ## Failure Modes - **过早收敛**:没问约束就选方案。 - **泛化方向**:方向像 grant topic,不像 paper idea。 - **忽视已有工作拥挤度**:热门方向 novelty 风险很高。 - **实验不可行**:idea 好听但用户资源无法验证。 ## Handoff - 如果用户要多个风险/成本版本 → `generate-idea-variants` - 如果候选 idea 需要投稿前风险扫描 → `identify-fatal-concerns` - 如果要把 idea 转成实验闭环 → `map-claims-to-experiments` - 如果方向已定,需要写 intro 论证链 → `build-intro-argument-chain` ## Example Input: ``` user_idea_seed: 我想做一个关于 LLM agent memory 的 idea,但不知道怎么创新。 constraints: 只能用 7B 以下模型,2 周实验。 ``` Output: ``` clarifying_questions: 1. 你想优化 memory 的 representation、retrieval、update,还是 evaluation? 2. 目标 agent 是科研、代码、网页还是通用任务? 3. 你能否构造错误记忆或长期一致性测试? research_axes: - memory objective: recall-oriented vs argument-state-oriented - update policy: append-only vs contradiction-aware - evaluation: recall accuracy vs decision/evidence consistency current_risk_reading: novelty: high, memory-augmented agent 已很拥挤。 evaluation: high, 需要避免只做 toy recall。 candidate_directions: D1: argument-state memory for research agents. insight: 科研 agent 更需要维护 claim-decision-evidence 一致性,而不只是 recall。 method sketch: memory slot 存储 claim/evidence/status 三元组。 test: 构造跨轮 claim consistency benchmark。 recommended_next_step: 先做 D1 的低成本诊断 benchmark,确认现有 memory 系统是否真的在 argument consistency 上失败。 ```