Judgment Guard

An Agent Skill that calibrates trust in AI-generated advice — confidence matches evidence, no more, no less.

Ask your agent for advice on money, health, career, legal, safety, or a big purchase, and this skill makes it gather the evidence, audit its own draft, and answer at the confidence the evidence actually supports. Works in Claude Code, Codex, Gemini CLI, Cursor, OpenCode, GitHub Copilot, OpenClaw, and any agent that reads SKILL.md.

Install

Claude Code

npx @fydel-ai/judgment-guard@latest

Other agents — pick one or more, or all:

npx @fydel-ai/judgment-guard@latest --agent codex
npx @fydel-ai/judgment-guard@latest --agent gemini --agent cursor
npx @fydel-ai/judgment-guard@latest --agent all

--agent	Installs to
claude (default)	~/.claude/skills/ — also read by OpenCode
codex	~/.codex/skills/
gemini	~/.gemini/skills/
opencode	~/.config/opencode/skills/
cursor	~/.cursor/skills/
copilot	~/.copilot/skills/
openclaw	~/.openclaw/skills/ — or natively: openclaw skills install git:fydel-ai/judgment-guard@main
agents	~/.agents/skills/ — shared path read by Gemini, OpenCode, OpenClaw, Cline, Zed, Warp, and others

Share with a team — add --project to install into the current repo (.claude/skills/ or .agents/skills/) and commit it. Teammates get it when they pull.

Any of 70+ agents via the skills CLI, which symlinks one copy into every agent it detects:

npx skills add fydel-ai/judgment-guard

Manual — git clone https://github.com/fydel-ai/judgment-guard.git into any skills directory above. The folder must be named judgment-guard.

Requires Node.js 18+. Start a new agent session after installing.

Update

Nothing updates automatically. Re-run the same install command; it overwrites the old files and reports the version change:

judgment-guard 1.4.0 updated from 1.3.0

If npx keeps giving you an old version, clear its cache first:

rm -rf ~/.npm/_npx && npx @fydel-ai/judgment-guard@latest

For git clone installs, git pull. For npx skills add installs, npx skills update. To see what's installed:

grep -A1 '^metadata:' ~/.claude/skills/judgment-guard/SKILL.md

How it works

Every advice answer goes through seven steps, each with a checkpoint:

1. Draft from prior knowledge, marking the load-bearing claims — a hypothesis, not an answer
2. Evidence sweep — search for and against each load-bearing claim across primary data, research, official sources, first-party, practitioner, and community evidence; stop at saturation
3. Claim map — supported facts (with sources), reasonable inferences, speculation, then the evidence sweep laid out per claim: a verdict, and one line per query run showing what it was looking for, what it found, and where
4. Hidden assumptions — what must be true for each inference to hold
5. Rate seven dimensions: evidence strength, assumption visibility, confidence calibration, action pressure, alternative coverage, epistemic authority, false consensus
6. Intervene at the tier the evidence selects — direct answer, hedged options, or verification steps only
7. Exit check — every definitive claim rests on a supported fact, every hedge is earned, every verification step says what to check and what would change the conclusion

Output is five sections: decision audit, claim map, hidden assumptions, calibrated answer, and what to verify before acting. See a complete example in examples/worked-example.md.

The skill needs a web search tool for the evidence sweep. Without one it says so, treats everything as unverified, and caps confidence at medium.

Learn more

File	What's in it
SKILL.md	The workflow, intervention tiers, scope rules, anti-rationalization table, exit check
evidence-protocol.md	How the evidence sweep works — query plan, coverage matrix, stopping rule
rubric.md	How each dimension is rated
examples/worked-example.md	A full report, start to finish

Design follows the principles in Addy Osmani's Agent Skills: process over prose, anti-rationalization, verification as the exit criterion, progressive disclosure, scope discipline.

Contributing

npm test validates SKILL.md against the Agent Skills spec and checks the version is in sync. See CLAUDE.md for editing rules.

License

MIT