01 ▸ Workflow Position
User Request (Vague)
│
▼
┌─────────────────────────────┐
│ clarification (ACTIVE) │
│ De-risk + Codebase Search │
└──────────────┬──────────────┘
│
┌────────┴────────┐
▼ ▼
Simple (5-8) Complex (10-15)
│ │
▼ ▼
Plan Crafting Milestone Planning
02 ▸ When to use?
"I want to...", "I need...", "Let's build..." — it's the starting point for every request where scope isn't immediately clear.
Key Triggers
- When request is vague and has multiple implementation paths
- When there's potential conflict with the existing codebase
- When the user hasn't fully defined what they want yet
03 ▸ How it works?
Core Principle: Explore first. Recon the codebase before asking, then ask confident questions grounded in the findings — iterating until ambiguity is gone.
Step-by-step Execution
- Identify Ambiguity: Spot unclear points in the request.
- Recon First + Informed Questions: Dispatch recon sub-agents before the first question, then ask rounds of evidence-grounded questions (bundle up to 4 independent ones).
- Synthesize Answer + Search Results: Cross-verify answers with code findings. If still vague, ask the next question.
- Generate Context Brief: Document goals, scope, constraints, success criteria, and complexity assessment.
- Auto-routing based on Complexity: Decide the next workflow based on the score (5-15).
04 ▸ Complexity Assessment
| Signal | Simple (1) | Medium (2) | Complex (3) |
|---|
| Scope | Single feature | 2-3 components | 4+ components |
| File Impact | ≤3 files | 4-8 files | 9+ files |
| Interface | Existing internal | Interface expansion | New interface def |
| Dependency | No order | Linear chain | Branching DAG |
| Risk Surface | No integration risk | Internal integration | External systems |
05 ▸ Hard Gates
Mandatory Rules
- Explore before you ask — never ask the user what the codebase can answer.
- Always search codebase via sub-agents — no assumptions or guesses.
- No implementation until scope is clear — no subjective "it's enough" judgments.
- Questions must narrow the scope — no repeating abstract questions.