adversarial-verify
Skill0 starsChanged 4 months agoMITView on GitHub
Use when findings, claims, or conclusions must be trusted before being reported or acted on — after an audit, code review, research sweep, or any agent-produced list of issues. Triggers: "verify these findings", "are these real", "kurie tikri", "patikrink radinius", "ar tikrai taip". Invoke proactively before presenting audit or review results to the user. Not for: raw discovery (agent-fanout), or verifying your own single change (universal-loop and verification-before-completion cover that).
7 other projects use a version of this.
Adopted by
Most recently taken up by
- shuymn/skillstheir own version, first seen Nov 1, 2026
- f0909172434/rigorgraphtheir own version, first seen Oct 11, 2026
- turntuptechnologies-ai/skillstheir own version, first seen Aug 6, 2026
- ulpi-io/autonomous-engineeringtheir own version, first seen Jul 19, 2026
- NUX-Design/claude-skills-fable-opustheir own version, first seen Jul 7, 2026
- Zavelinski/claude-code-adversarial-verifytheir own version, first seen Jun 26, 2026
Other versions7 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 7
- NUX-Design/claude-skills-fable-opus★ 24same name
- ulpi-io/autonomous-engineering★ 2same name
- f0909172434/rigorgraph★ 1same name
- jpoindexter/ai-agent-skills★ 1same name
- turntuptechnologies-ai/skills★ 1same name
- shuymn/skills★ 0same name
- Zavelinski/claude-code-adversarial-verify★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository9 other files it gives its agents.
- agent-fanoutSkill
- compounding-engineeringSkill
- deployment-checkSkill
- goal-contractSkill
- karpathy-guidelinesSkill
- recon-firstSkill
- two-strike-debugSkill
- universal-loopSkill
- writing-claude-mdSkill
The file
We split this skill into the rules it gives an agent
2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 6 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 34
If a claim can fail in several ways, N identical skeptics waste the budget.
- Distinctive: line 51
After verification, one final agent gets a single question: "What is missing — which angle wasn't searched, which claim wasn't verified, …
“Common” is measured: kept by at least 6 in 10 of the 6 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Adversarial Verify
Triggers automatically when relevant; invoke directly with /adversarial-verify or ask, e.g. "verify these findings — which are real?".
Plausible-but-wrong findings are the main failure mode of LLM review, and verification by confirmation cannot catch them — a confirmer shares the finder's biases. Refutation catches them. Silence is a feature: a short list of survivors beats a long list of maybes.
Core move — the skeptic panel
For each finding, spawn N independent skeptics whose task is to REFUTE it:
"Try to refute this claim:
<finding>. Look for reasons it is wrong, unreproducible, or misread. If uncertain, default to refuted=true."
Majority refute → the finding dies. Report only survivors, and say how many died — the kill count is evidence the filter actually ran.
Lens diversity beats redundancy
If a claim can fail in several ways, N identical skeptics waste the budget. Give each a different lens: correctness / security / performance / does-it-actually-reproduce. Diversity catches failure modes redundancy can't.
Rules
- Finder ≠ verifier. The verifying agent must be different from the one that produced the finding — self-review confirms its own assumptions.
- Default to refuted. A skeptic who "can't rule it out" has not confirmed anything.
- Scale to stakes: quick check → 1 skeptic per finding; audit going to a decision-maker → 3–5 lenses per finding.
- No silent caps: sampled? truncated? Say what was dropped.
Completeness critic — the last layer
After verification, one final agent gets a single question: "What is missing — which angle wasn't searched, which claim wasn't verified, which source wasn't read?" Its findings are the next work wave, not footnotes.
Anti-patterns
- ✗ Verifying by asking "is this correct?" — confirmers agree too easily.
- ✗ The same agent finding and verifying.
- ✗ Reporting all findings with confidence labels instead of killing the weak ones — the reader's time is the budget you're spending.
Relationship to other skills
- agent-fanout produces the findings this skill filters.
- verification-before-completion is the final gate for your OWN work; this skill is for claims produced by review at scale.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.