## 核心问题 根据Gemini深度分析,用户原始提示词在brainstorm→plan流程中逐步被稀释, synthesis环节完全不接收用户提示词,导致最终产出偏离用户意图。 ## 关键修改 ### 1. synthesis.md - 修复最严重偏差环节 - 新增从workflow-session.json加载原始用户提示词 - Agent流程Step 1: load_original_user_prompt (HIGHEST priority) - 新增"用户意图对齐"为首要合成要求 - 添加意图追溯和偏差警告完成标准 ### 2. concept-clarify.md - 添加意图验证 - 加载original_user_intent从session metadata - 新增"用户意图对齐"为最高优先级扫描类别 - 检查目标一致性、范围匹配、偏差检测 ### 3. action-plan-verify.md - 强化验证门禁 - workflow-session.json作为主要参考源 - 新增"用户意图对齐"为CRITICAL级别检测 - 违反用户意图标记为CRITICAL严重性 ### 4. plan.md - 确立意图优先级 - 明确用户原始意图始终为主要参考 - 优先级规则: 当前用户提示词 > synthesis > topic-framework ### 5. artifacts.md - 推广结构化格式 - 添加推荐的GOAL/SCOPE/CONTEXT结构化格式 - 强调用户意图保存的重要性 ### 6. auto-parallel.md - 编排器完整性 - 推荐结构化提示词格式 - Phase 1标注用户意图存储 - Phase 3明确synthesis加载用户意图为最高优先级 - Agent模板强调用户意图为PRIMARY reference ## 影响力提升 用户提示词影响力: 30% → 95% ## 参考 分析报告由Gemini生成,识别了5个关键问题点并提供4条改进建议 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
12 KiB
name, description, argument-hint, allowed-tools
| name | description | argument-hint | allowed-tools |
|---|---|---|---|
| concept-clarify | Identify underspecified areas in brainstorming artifacts through targeted clarification questions before action planning | [optional: --session session-id] | Read(*), Write(*), Edit(*), TodoWrite(*), Glob(*), Bash(*) |
User Input
$ARGUMENTS
You MUST consider the user input before proceeding (if not empty).
Outline
Goal: Detect and reduce ambiguity or missing decision points in planning artifacts before moving to task generation. Supports both brainstorm and plan workflows.
Timing:
- Brainstorm mode: Runs AFTER
/workflow:brainstorm:synthesisand BEFORE/workflow:plan - Plan mode: Runs AFTER Phase 3 (concept-enhanced) and BEFORE Phase 4 (task-generate) within
/workflow:plan
This serves as a quality gate to ensure conceptual clarity before detailed task planning or generation.
Execution steps:
-
Session Detection & Validation
# Detect active workflow session IF --session parameter provided: session_id = provided session ELSE: CHECK: .workflow/.active-* marker files IF active_session EXISTS: session_id = get_active_session() ELSE: ERROR: "No active workflow session found. Use --session <session-id> or start a session." EXIT # Mode detection: plan vs brainstorm brainstorm_dir = .workflow/WFS-{session}/.brainstorming/ process_dir = .workflow/WFS-{session}/.process/ IF EXISTS(process_dir/ANALYSIS_RESULTS.md): clarify_mode = "plan" primary_artifact = process_dir/ANALYSIS_RESULTS.md INFO: "Plan mode: Analyzing ANALYSIS_RESULTS.md" ELSE IF EXISTS(brainstorm_dir/synthesis-specification.md): clarify_mode = "brainstorm" primary_artifact = brainstorm_dir/synthesis-specification.md INFO: "Brainstorm mode: Analyzing synthesis-specification.md" ELSE: ERROR: "No valid artifact found. Run /workflow:brainstorm:synthesis or /workflow:plan first" EXIT # Mode-specific validation IF clarify_mode == "brainstorm": CHECK: brainstorm_dir/topic-framework.md IF NOT EXISTS: WARN: "topic-framework.md not found. Verification will be limited." -
Load Artifacts (Mode-Aware)
# Load primary artifact (determined in step 1) primary_content = Read(primary_artifact) # Load user's original intent from session metadata session_metadata = Read(.workflow/WFS-{session}/workflow-session.json) original_user_intent = session_metadata.project || session_metadata.description IF NOT original_user_intent: WARN: "No original user intent found in session metadata" original_user_intent = "Not available" # Load mode-specific supplementary artifacts IF clarify_mode == "brainstorm": topic_framework = Read(brainstorm_dir + "/topic-framework.md") # if exists role_analyses = Glob(brainstorm_dir + "/*/analysis.md") participating_roles = extract_role_names(role_analyses) # Plan mode: primary_content (ANALYSIS_RESULTS.md) is self-contained -
Ambiguity & Coverage Scan
Perform structured scan using this taxonomy. For each category, mark status: Clear / Partial / Missing.
⚠️ User Intent Alignment (NEW - HIGHEST PRIORITY):
- Primary artifact alignment with original user intent
- Goal consistency between artifact and user's stated objectives
- Scope match with user's requirements
- Success criteria reflects user's expectations
- Any unexplained deviations from user intent
Requirements Clarity:
- Functional requirements specificity and measurability
- Non-functional requirements with quantified targets
- Business requirements with success metrics
- Acceptance criteria completeness
Architecture & Design Clarity:
- Architecture decisions with rationale
- Data model completeness (entities, relationships, constraints)
- Technology stack justification
- Integration points and API contracts
User Experience & Interface:
- User journey completeness
- Critical interaction flows
- Error/edge case handling
- Accessibility and localization considerations
Implementation Feasibility:
- Team capability vs. required skills
- External dependencies and failure modes
- Resource constraints (timeline, personnel)
- Technical constraints and tradeoffs
Risk & Mitigation:
- Critical risks identified
- Mitigation strategies defined
- Success factors clarity
- Monitoring and quality gates
Process & Collaboration:
- Role responsibilities and handoffs
- Collaboration patterns defined
- Timeline and milestone clarity
- Dependency management strategy
Decision Traceability:
- Controversial points documented
- Alternatives considered and rejected
- Decision rationale clarity
- Consensus vs. dissent tracking
Terminology & Consistency:
- Canonical terms defined
- Consistent naming across artifacts
- No unresolved placeholders (TODO, TBD, ???)
For each category with Partial or Missing status, add to candidate question queue unless:
- Clarification would not materially change implementation strategy
- Information is better deferred to planning phase
-
Generate Prioritized Question Queue
Internally generate prioritized queue of candidate questions (maximum 5):
Constraints:
- Maximum 5 questions per session
- Each question must be answerable with:
- Multiple-choice (2-5 mutually exclusive options), OR
- Short answer (≤5 words)
- Only include questions whose answers materially impact:
- Architecture decisions
- Data modeling
- Task decomposition
- Risk mitigation
- Success criteria
- Ensure category coverage balance
- Favor clarifications that reduce downstream rework risk
Prioritization Heuristic:
priority_score = (impact_on_planning * 0.4) + (uncertainty_level * 0.3) + (risk_if_unresolved * 0.3)If zero high-impact ambiguities found, proceed to Step 8 (report success).
-
Sequential Question Loop (Interactive)
Present EXACTLY ONE question at a time:
Multiple-choice format:
**Question {N}/5**: {Question text} | Option | Description | |--------|-------------| | A | {Option A description} | | B | {Option B description} | | C | {Option C description} | | D | {Option D description} | | Short | Provide different answer (≤5 words) |Short-answer format:
**Question {N}/5**: {Question text} Format: Short answer (≤5 words)Answer Validation:
- Validate answer maps to option or fits ≤5 word constraint
- If ambiguous, ask quick disambiguation (doesn't count as new question)
- Once satisfactory, record in working memory and proceed to next question
Stop Conditions:
- All critical ambiguities resolved
- User signals completion ("done", "no more", "proceed")
- Reached 5 questions
Never reveal future queued questions in advance.
-
Integration After Each Answer (Incremental Update)
After each accepted answer:
# Ensure Clarifications section exists IF primary_content NOT contains "## Clarifications": Insert "## Clarifications" section after first heading # Create session subsection IF NOT contains "### Session YYYY-MM-DD": Create "### Session {today's date}" under "## Clarifications" # Append clarification entry APPEND: "- Q: {question} → A: {answer}" # Apply clarification to appropriate section CASE category: Functional Requirements → Update "## Requirements" or equivalent section Architecture → Update "## Architecture" or "## Design" sections User Experience → Update "## UI/UX" or "## User Experience" sections Risk → Update "## Risks" or "## Risk Assessment" sections Process → Update "## Process" or "## Implementation" sections Data Model → Update "## Data Model" or "## Database" sections Non-Functional → Update "## Non-Functional Requirements" or equivalent # Remove obsolete/contradictory statements IF clarification invalidates existing statement: Replace statement instead of duplicating # Save immediately to primary_artifact Write(primary_artifact) -
Validation After Each Write
- Clarifications section contains exactly one bullet per accepted answer
- Total asked questions ≤ 5
- Updated sections contain no lingering placeholders
- No contradictory earlier statements remain
- Markdown structure valid
- Terminology consistent across all updated sections
-
Completion Report
After questioning loop ends or early termination:
## ✅ Concept Verification Complete **Session**: WFS-{session-id} **Mode**: {clarify_mode} **Questions Asked**: {count}/5 **Artifacts Updated**: {primary_artifact filename} **Sections Touched**: {list section names} ### Coverage Summary | Category | Status | Notes | |----------|--------|-------| | Requirements Clarity | ✅ Resolved | Acceptance criteria quantified | | Architecture & Design | ✅ Clear | No ambiguities found | | Implementation Feasibility | ⚠️ Deferred | Team training plan to be defined in IMPL_PLAN | | Risk & Mitigation | ✅ Resolved | Critical risks now have mitigation strategies | | ... | ... | ... | **Legend**: - ✅ Resolved: Was Partial/Missing, now addressed - ✅ Clear: Already sufficient - ⚠️ Deferred: Low impact, better suited for planning phase - ❌ Outstanding: Still Partial/Missing but question quota reached ### Recommendations - ✅ **PROCEED to /workflow:plan**: Conceptual foundation is clear - OR ⚠️ **Address Outstanding Items First**: {list critical outstanding items} - OR 🔄 **Run /workflow:concept-clarify Again**: If new information available ### Next Steps **If Outstanding Items Exist**: 1. TodoWrite tracking for outstanding issues 2. Call conceptual-planning-agent to fix conceptual gaps 3. Wait for user confirmation before proceeding **If No Issues**: ```bash /workflow:plan # Generate IMPL_PLAN.md and task.json -
Update Session Metadata
# Update metadata based on mode IF clarify_mode == "brainstorm": phase_key = "BRAINSTORM" ELSE: # plan mode phase_key = "PLAN" # Update session metadata { "phases": { "{phase_key}": { "status": "concept_verified", "concept_verification": { "completed": true, "completed_at": "timestamp", "mode": "{clarify_mode}", "questions_asked": {count}, "categories_clarified": [{list}], "outstanding_items": [], "recommendation": "PROCEED" # or "ADDRESS_OUTSTANDING" } } } }
Behavior Rules
- If no meaningful ambiguities found: Report "No critical ambiguities detected. Conceptual foundation is clear." and suggest proceeding to
/workflow:plan. - If synthesis-specification.md missing: Instruct user to run
/workflow:brainstorm:synthesisfirst. - Never exceed 5 questions (disambiguation retries don't count as new questions).
- Respect user early termination: Signals like "stop", "done", "proceed" should stop questioning.
- If quota reached with high-impact items unresolved: Explicitly flag them under "Outstanding" with recommendation to address before planning.
- Avoid speculative tech stack questions unless absence blocks conceptual clarity.
Operating Principles
Context Efficiency
- Minimal high-signal tokens: Focus on actionable clarifications
- Progressive disclosure: Load artifacts incrementally
- Deterministic results: Rerunning without changes produces consistent analysis
Verification Guidelines
- NEVER hallucinate missing sections: Report them accurately
- Prioritize high-impact ambiguities: Focus on what affects planning
- Use examples over exhaustive rules: Cite specific instances
- Report zero issues gracefully: Emit success report with coverage statistics
- Update incrementally: Save after each answer to minimize context loss
Context
{ARGS}