Judgment Guard

A Claude Code skill that calibrates trust in AI-generated advice — ensuring confidence matches evidence, no more, no less.

When you ask Claude for advice, recommendations, comparisons, or prioritization on consequential topics — money, health, career, legal, safety, purchases — this skill checks whether evidence quality, recency, or uncertainty should change the confidence of the answer, then rewrites it to match.

What it does

1. Evaluates draft answers on seven dimensions: evidence strength, assumption visibility, confidence calibration, action pressure, alternative coverage, epistemic authority, and false consensus
2. Applies an intervention policy:
   • Strong evidence — direct, confident answer with assumptions surfaced inline
   • Mixed evidence — multiple interpretations, counterpoints, and what would change the conclusion
   • Weak evidence — no ranked recommendations, only possibilities and verification steps
3. Outputs a structured report: decision audit, claim map, hidden assumptions, calibrated answer, and verification checklist

Install

Requirements: Claude Code and Node.js 18+

Quick install (recommended)

Install judgment-guard: run npx @fydel-ai/judgment-guard

Manual install

Clone the repo into your user skills directory:

Install judgment-guard: run git clone https://github.com/fydel-ai/judgment-guard.git ~/.claude/skills/judgment-guard

Either method creates:

~/.claude/skills/judgment-guard/SKILL.md

Project install

To share the skill with teammates in a repo, copy it into the project skills directory:

Add judgment-guard to this project: run cp -Rf ~/.claude/skills/judgment-guard .claude/skills/judgment-guard && rm -rf .claude/skills/judgment-guard/.git

Committed to a repository, this gives teammates:

.claude/skills/judgment-guard/SKILL.md

Usage

Claude Code can load the skill automatically when the prompt calls for advice on consequential topics, or you can invoke it directly with /judgment-guard.

Files

| File | Purpose | |------|---------| | SKILL.md | Skill definition — trigger conditions, evaluation dimensions, intervention policy, output format | | rubric.md | Scoring criteria for all seven evaluation dimensions | | templates/report_template.md | Blank template for the five output sections |

Example output

### Decision audit
- Evidence strength: medium
- Assumption visibility: low
- Confidence calibration: overstated
- Action pressure: high
- Alternative coverage: low
- Epistemic authority: appropriate
- False consensus: none detected

### Claim map
**Supported facts**
- ...

**Reasonable inferences**
- ...

**Speculation / missing evidence**
- ...

### Hidden assumptions
**What must be true for the inferences above to hold**
- ...

**Assumptions the user may not have considered**
- ...

### Calibrated answer
(rewritten response with confidence matched to evidence)

### What to verify before acting
1. ...
2. ...
3. ...

License

MIT