Confidence-Weighted Research Skill

Confidence-Weighted Research

Research outputs with explicit confidence levels and source hierarchy. Halts hallucination propagation by tagging unverified claims.

Core principle: Every claim tagged with confidence level and source. Investigation, not confirmation.

Source Hierarchy

| Priority | Source | Confidence | Examples | |----------|--------|------------|----------| | 1 | Context7 | HIGH | Official library docs, version-specific | | 2 | Official Docs | HIGH | Documentation, API references | | 3 | WebSearch (verified) | MEDIUM | Community patterns, verified claims | | 4 | WebSearch (unverified) | LOW | Single sources, training data only | | 5 | Inference | LOW | LLM-generated without sources |

Research Output Format

# Research: [Topic]

## Executive Summary
[2-3 sentence overview]

## Findings

### Finding 1: [Title]
**Claim:** [What was found]
**Confidence:** HIGH | MEDIUM | LOW
**Source:** context7 | official-docs | websearch | inference
**Verified:** YES | NO | PARTIAL
**Evidence:**
- [Source 1 with URL]
- [Source 2 with URL]

### Finding 2: [Title]
...

Verification Protocol

For each finding:

Is this from Context7 or official docs? → HIGH confidence
Is this from verified WebSearch? → MEDIUM confidence
Is this from unverified source? → LOW confidence, flag
Multiple sources agree? → Increase confidence one level

Anti-Pattern: Confirmation Bias

Bad research: Start with hypothesis, find supporting evidence

Good research: Gather evidence, form conclusions from evidence

Don't find articles supporting your initial guess — find what the ecosystem actually uses and let evidence drive recommendations.

Confidence Levels

| Level | Meaning | When to Use | |-------|---------|--------------| | HIGH | Verified by authoritative source | State as fact | | MEDIUM | Verified by community, unverified by authority | State with attribution | | LOW | Single source or inference only | Flag as needing validation |

Research Modes

| Mode | Trigger | Scope | |------|---------|-------| | Ecosystem | What exists for X? | Libraries, frameworks, standard stack | | Feasibility | Can we do X? | Technical achievability, constraints | | Comparison | Compare A vs B | Features, performance, DX |

Tool Strategy

Priority Order

Context7 (highest) — Authoritative, current, version-aware
Official Docs via WebFetch — Authoritative sources
WebSearch — Ecosystem discovery
Inference (lowest) — Only when other sources unavailable

For Each Finding

Is Context7 available? → YES → HIGH confidence
        ↓ NO
Is official docs available? → YES → HIGH confidence
        ↓ NO
Are multiple credible sources available? → YES → MEDIUM confidence
        ↓ NO
Single source or inference? → LOW confidence, flag for validation

Pre-Research Checklist

[ ] Scan Context7 for library documentation
[ ] Check official docs for authoritative sources
[ ] Use WebSearch for community patterns
[ ] Verify claims against multiple sources

Post-Research Checklist

[ ] All findings tagged with confidence level
[ ] All findings have source attribution
[ ] LOW confidence findings flagged for validation
[ ] Negative claims verified with official docs

Integration

Use with:

/ask:team — For knowledge synthesis
/plan:team — For architecture decisions
/research commands — For domain research

Anti-Hallucination Discipline

Verify before asserting — Check Context7 or official docs
Prefer current sources — Context7 and official docs trump training data
Flag uncertainty — LOW confidence when only training data
"I couldn't find X" is valuable — Investigate differently
Sources contradict — Surface the ambiguity

Agent Skills: Confidence-Weighted Research

Install this agent skill to your local

Skill Files