confidence-aware-planning

name	confidence-aware-planning
description	Plan differently when the agent has low certainty — ask for clarification or narrow scope instead of pretending full understanding.
origin	FlowDeck

Confidence-Aware Planning

Not every task comes with complete information. This skill enforces honest uncertainty signaling and adaptive planning when confidence is low.

Confidence Levels

Level	Meaning	Action
HIGH (≥80%)	Well-understood scope, clear precedent in codebase	Proceed to plan normally
MEDIUM (40–79%)	Partial understanding, some unknowns	Surface assumptions, narrow scope, flag for review
LOW (<40%)	Significant unknowns, no clear precedent	Ask clarifying questions first, do not plan until answered

Signals That Lower Confidence

• Codebase section not covered in .codebase/ARCHITECTURE.md
• No prior task.md for this feature area
• Request touches 5+ files with unclear dependencies
• Request uses domain jargon that doesn't appear in codebase
• No test coverage in the affected area (no test files found)

Workflow

Before planning ANY task:

1. Read relevant codebase docs (ARCHITECTURE.md, STACK.md, CONVENTIONS.md)
2. Scan affected files for context
3. Estimate confidence level
4. Act based on level:
   • HIGH: proceed to /fd-review
   • MEDIUM: write explicit assumptions at the top of PLAN.md, flag 3 highest risks
   • LOW: stop, ask clarifying questions, do not write PLAN.md until answered

Clarifying Question Format

When confidence is LOW, ask in this format:

Before I can plan this, I need to understand:

1. [Question about scope/behavior]
2. [Question about constraint or requirement]
3. [Question about existing system behavior]

I have LOW confidence because: [specific reason]

Assumption Declaration Format

When confidence is MEDIUM, include at the top of every plan:

## Assumptions (MEDIUM confidence)
- A1: [assumption] — if wrong, [consequence]
- A2: [assumption] — if wrong, [consequence]

## Risks
1. [risk]: [mitigation]

Non-negotiable

Never write a plan that pretends HIGH confidence when the agent actually has LOW confidence. False certainty leads to wrong implementations and wasted effort.