confidence-check
Use before implementing when uncertainty exists. Weighted scoring across 5 checks (requires ≥80%). Triggers: "before implementing", "verify readiness", "should I proceed", "am I ready". If thinking "this is overkill" - use it.
When & Why to Use This Skill
The Confidence Check skill is a high-impact pre-implementation gate designed to validate technical strategies before code execution. By enforcing a weighted scoring system across five critical checks—duplicate detection, architecture compliance, documentation verification, OSS referencing, and root cause identification—it prevents 'wrong-direction' development. This skill is essential for optimizing developer productivity and minimizing token waste by ensuring that every implementation is necessary, architecturally sound, and well-researched.
Use Cases
- Feature Planning: Use before starting any new feature to ensure no similar implementation already exists, preventing codebase bloat and technical debt.
- Architectural Alignment: Automatically verify that proposed refactors or new modules strictly adhere to the project's established tech stack and patterns defined in CLAUDE.md.
- Bug Fix Validation: Ensure the root cause of an issue is fully identified and documented before modifying production code, reducing the risk of regression.
- Agentic Guardrails: Set a mandatory confidence threshold (80%) for AI agents to stop and request clarification when requirements are vague or implementation risks are high.
| name | confidence-check |
|---|---|
| description | Use before implementing when uncertainty exists. Weighted scoring across 5 checks (requires ≥80%). Triggers: "before implementing", "verify readiness", "should I proceed", "am I ready". If thinking "this is overkill" - use it. |
Confidence Check
Pre-implementation gate. Spend 100-200 tokens here to save 5,000-50,000 tokens on wrong-direction work.
When to Use
USE PROACTIVELY - before EVERY implementation, not just when uncertain:
- Starting any feature, fix, or refactor
- About to write production code
- Architecture/stack decisions to make
- Integrating with unfamiliar code
Especially critical when:
- Stack is new or unfamiliar
- Requirements feel vague
- Codebase is large or complex
- User seems confident (overconfidence = highest risk)
DO NOT use for:
- Pure research/exploration tasks
- Reading/explaining existing code
- Documentation-only changes
MANDATORY FIRST STEP
TodoWrite: Create 5 items (1 per check)
- Search for duplicate implementations (Grep/Glob)
- Verify architecture compliance (CLAUDE.md, patterns)
- Check official documentation (Context7/WebFetch)
- Find working OSS reference (Tavily/WebSearch)
- Identify root cause (errors, logs, traces)
5 Checks (Weighted)
| Check | Weight | Pass Criteria |
|---|---|---|
| No Duplicates | 25% | No similar implementation exists |
| Architecture Compliant | 25% | Uses existing stack/patterns |
| Official Docs Verified | 20% | Official docs reviewed |
| Working OSS Reference | 15% | Proven implementation found |
| Root Cause Identified | 15% | Root cause clear |
Task-specific variants:
- Bug Fix: Root cause (40%) + Docs (30%) + OSS (30%)
- New Feature: Duplicates (40%) + Architecture (30%) + Docs (30%)
- Refactor: Architecture (50%) + Duplicates (30%) + OSS (20%)
Decision Thresholds
| Score | Action |
|---|---|
| ≥80% | ✅ Proceed to implementation |
| 70-79% | ⚠️ Present alternatives, ask clarifying questions |
| <70% | ❌ STOP - Request more context from user |
Output Format
Confidence Checks:
[✅/❌] No duplicate implementations found
[✅/❌] Uses existing tech stack
[✅/❌] Official documentation verified
[✅/❌] Working OSS implementation found
[✅/❌] Root cause identified
Confidence: X.XX (XX%)
Decision: [Proceed/Ask Questions/Stop]
Response Templates
"Skip the check, this is straightforward"
Straightforward tasks fail 40% of the time from duplicate implementations or architecture mismatches. Confidence check takes 2 minutes. Which checks did you already complete?
Red Flags
| Thought | Reality |
|---|---|
| "This is too simple to check" | 40% of "simple" tasks duplicate existing code |
| "I already know the architecture" | Assumptions cause 30% of rework |
| "Official docs take too long" | 5 min reading saves 3 hours debugging |