adversarial-review

Have an independent model try to break a proposed change and return its findings as testable hypotheses, each with the smallest experiment that would refute it.

Skill by whichguy

IndexedMIT★ 1

Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/whichguy-adversarial-review

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

About

Have an independent model try to break a proposed change and return its findings as testable hypotheses, each with the smallest experiment that would refute it. Use before shipping any prompt, review-focus, card or harness change decided from experiments; for "adversarial review", "red-team this change", "what could this break". Writes findings.json for rubric-eval.

Use cases: agent-building

Reviews

Write a review

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.

Tools: No tools found
Secrets: None asked for
Install scripts: None found
Runs as: Python package

Network hosts it talks to 1

codeload.github.com · found in install command

Files it touches

~/.claude/skills/adversarial-review

What's inside

Skill name: adversarial-review
Description: Have an independent model try to break a proposed change and return its findings as testable hypotheses, each with the smallest experiment that would refute it. Use before shipping any prompt, review-focus, card or harness change decided from experiments; for "adversarial review", "red-team this change", "what could this break". Writes findings.json for rubric-eval.
License: MIT
Instructions: 25 lines in SKILL.md · view on GitHub

Before installing

Ships executable scripts

Files (5)

SKILL.md
SPEC.md
findings.schema.json · references/
adversarial-review · scripts/
adversarial_review.py · scripts/

SKILL.md (first lines)

# adversarial-review

Read [SPEC.md](SPEC.md) for the findings contract and how findings become
experiments. It composes the `rubric-eval` skill's model call, so install both.

```sh
A=scripts/adversarial-review
$A review --name "prune review" --change change.md --evidence evidence.md --out RUN/adversarial --reviewer luna
$A validate RUN/adversarial/findings.json
```

- `change.md`: the exact text before and after, and where it ships.
- `evidence.md`: the experiment results the change rests on (tables, arms,
  subject and judge models, conditions).
- `--reviewer`: a model alias or `host:model@effort` spec (see rubric-eval's
  `model_call.py`) from a family other than the one that proposed or produced
  the change; default `luna` (Codex gpt-6-luna, max effort).

Then, for each testable high or medium finding, run its experiment with
`rubric-eval` and record whether the stated refuting result occurred. A finding
that holds sends the change back for revision; one that is refuted is recorded
with its evidence. Findings marked not testable are decided by judgement, with
the reason written down.

From whichguy/skill-craft, MIT.

Signals

Stars: 1
Forks: 0
Open issues: 0
Last commit (this folder): Oct 5, 2026 (24 days ago)
Last release: Aug 15, 2026 (2 months ago)

From GitHub (fetched Oct 24, 2026).

Details

License: MIT
Repository: github.com/whichguy/skill-craft
First seen: Oct 24, 2026
Last checked: Oct 24, 2026
Source: GitHub github-graphql

More from whichguy

Cite this page

adversarial-review skill. Agent Mart. https://agentmart.dev/skills/whichguy-adversarial-review (last updated 2026-10-24)