shield_lock
金大哥
← 返回 Skill 列表
extension
分类: AI Agent 能力无需 API Key

do-discuss

在通过苏格拉底式对话进行规划之前，构建结构化的问题框架。当问题模糊、不明确或过于宽泛而无法直接规划时使用。当用户描述的是症状、感受或情况而非具体需求时也应触发——例如“我不确定实际问题是什么”、“帮我思考这个问题”、“我应该考虑什么”或“这种情况一直发生但我不知道为什么”。此外，在头脑风暴产生需要框架的选定方向时，或者do-debug揭示具有架构范围的根本原因时也应触发。当需求已经准备好可以规划（直接进入do-plan）、用户报告了一个可重现的缺陷（使用do-debug）或没有特定问题的创意构思（使用头脑风暴）时不要使用。

person作者: jakexiaohubgithub
open_in_new仓库download下载资源包

Frame the actual problem through tiered Socratic dialogue before planning: intake → clarify → investigate → explore → frame → handoff.

Investigate and explore are conditional — triggered by tier escalation, not always run.

Tiered Interaction Model

| Tier | Mode | Trigger | Agents | |------|------|---------|--------| | 1 | Socratic dialogue (main thread) | Always starts here | None | | 2 | Codebase-assisted | Codebase-dependent unknown blocks framing AND user cannot answer | @scout or @researcher | | 3 | Multi-perspective exploration | Ambiguous scope + 2+ one-way-door decisions after clarify + investigate | @framer team |

Phases

Subagent dispatch policy: Each role uses its specialized agent type. Every dispatch prompt MUST include:

  • The exact output file path (.scratch/<session>/<prescribed-filename>.md)
  • The constraint: "Write your complete output to that path. You may read any repository file. Do NOT edit, create, or delete files outside .scratch/<session>/. Do NOT run build commands, tests, or destructive shell commands."

| Phase | Agent type | Rationale | |-------|-----------|-----------| | Investigate | @scout / @researcher | Breadth-first orientation or deep evidence gathering | | Explore | @framer | Perspective-committed, advisory, peer-reactive |

Session ID: Generate per SPINE.md Sessions convention. Carry forward into do-plan if user proceeds. Append to the session log at phase boundaries and tier escalations. All output paths below use <session> as placeholder.

1. Intake

Accept raw input, classify, redirect if wrong tool.

| Input shape | Action | |-------------|--------| | Has tasks, criteria, scope → plan-ready | Redirect to do-plan | | Reproducible defect with steps | Redirect to do-debug | | Pure ideation without a problem | Redirect to brainstorming | | < 1 sentence of problem context | Ask grounding question: "What situation are you trying to change?" | | Vague, ambiguous, or symptom-based | Proceed to clarify |

Detect upstream handoff context: brainstorming (selected direction) or do-debug (root cause confirmed, scope exceeds bug fix). When present, seed the known inventory from the upstream artifact.

2. Clarify

Socratic dialogue. Main thread only (tier 1). No subagent dispatch.

Derive the frame question — the single question whose answer unblocks planning. A good frame question is specific (names the affected system or behavior), answerable (finite set of possible answers), and scoped (answering it directly enables planning). Example: "Is the auth retry failure a client-side timeout or a server-side rate limit?"

  • Identify ambiguity buckets: goal, scope, constraints, stakeholder
  • Batch 2-4 contextually independent questions per round
  • Ask sequentially dependent questions one at a time
  • Track known / unknown inventory across rounds
  • 3-round budget
  • After each round: update inventory, check if frame question is answerable

Escalation trigger to tier 2: a codebase-dependent unknown blocks framing AND the user says "check the code" / "I don't know how it works" / cannot answer from domain knowledge.

3. Investigate (conditional, tier 2)

Subagent dispatch for codebase evidence when user cannot resolve unknowns.

| Agent | Use when | Output | |-------|----------|--------| | @scout | Orientation: "does X exist? where? what shape?" | .scratch/<session>/discuss-investigate-scout.md | | @researcher | Depth: "how does X work? what are the side effects?" | .scratch/<session>/discuss-investigate-researcher.md |

Dispatch context (include in every investigate prompt):

  • The specific unknown(s) to resolve (from the unknown inventory)

  • Current known / unknown inventory state

  • Why the user could not answer (domain gap vs. codebase gap)

  • Max 2 dispatch rounds, max 2 concurrent agents per round

  • Main thread synthesizes dispatch outputs into codebase_signals

  • After synthesis: return to clarify to re-check if the frame question is now answerable

Escalation trigger to tier 3: ambiguous scope AND key_decisions has 2+ one-way-door options after investigation.

4. Explore (conditional, tier 3)

Multi-perspective agent team when one-way-door decisions resist convergence.

  • Team name: discuss-explore
  • Three @framer personas:

| Role | Output | |------|--------| | stakeholder-advocate | .scratch/<session>/discuss-explore-stakeholder-advocate.md | | systems-thinker | .scratch/<session>/discuss-explore-systems-thinker.md | | skeptic | .scratch/<session>/discuss-explore-skeptic.md |

Dispatch context (include in every framer prompt):

  • Current problem_frame state (in-progress)
  • Full known / unknown inventory
  • Assigned perspective name
  • All three output paths (so the framer can read peer outputs)

Sequencing: dispatch all three in parallel → wait for completion → re-invoke each framer to read peer outputs and append a ## Peer Reactions section → main thread synthesizes. Irreconcilable positions become key_decisions in the frame.

5. Frame

Main thread only. Produce problem_frame artifact.

Write structured artifact to .scratch/<session>/discuss-frame.md using the template from references/frame-template.md.

The frame must be self-sufficient: understandable without chat history, all terms defined, no conversation references.

6. Handoff

Main thread only. Sole handoff authority.

Emit confidence-gated recommendation.

| Confidence | Declaration | |------------|-------------| | high | "Discussion complete. Proceed with /do-plan." | | medium | "Discussion complete with open assumptions. Proceed with /do-plan or resolve assumptions first." | | low | "Problem needs further exploration. Consider /brainstorming to clarify direction." |

Emit discussion_learnings proposals (never auto-applied). STOP. Do not invoke do-plan automatically. User decides next step.

Relationship to Adjacent Skills

do-discuss -> do-plan -> do-execute    (frame -> plan -> build)
brainstorming -> do-plan               (ideate -> plan)
do-discuss -> brainstorming            (when confidence is low, discuss recommends ideation)
do-debug -> do-discuss                 (when root cause reveals architectural scope)
  • do-discuss: "What is the actual problem?" (convergent, diagnostic)
  • brainstorming: "What should we build?" (divergent, creative)

Ask Policy

Ask whenever ambiguity materially affects the frame question. Two pressure points:

  • During clarify: when the user's answer reveals the problem is substantially different from what was initially described
  • Before handoff: when blocking unknowns remain that only the user can resolve

Never carry unresolved blocking unknowns silently into the frame artifact.

Termination Conditions

| Condition | Outcome | |-----------|---------| | Frame question answered + blocking unknowns resolved | complete | | User says "just plan it" + blocking unknowns acknowledged | complete with accepted risk | | 5 iteration cap reached (one iteration = one pass from clarify through frame) | Freeze best state, surface gaps | | Input classified as wrong tool | Redirect, no artifact |

Anti-Patterns

  • Asking leading questions that embed solutions
  • Proposing implementations during discuss (this is framing, not planning)
  • Framing the frame question as a solution choice ("should we use Redis?") instead of a diagnostic question ("is the bottleneck in storage or retrieval?")
  • Dispatching tier-2 agents for questions answerable by the user
  • Dispatching tier-3 team for simple problems with clear scope
  • Silently dropping blocking unknowns when user wants to proceed
  • Running discuss when input is already plan-ready
  • Producing a frame artifact that references conversation turns (violates self-sufficiency)