squad-review — Claude Code skill: what it does, how to install
squad-review runs six specialist code reviews on a branch diff, then tests every reported issue with an adversarial verifier.
MIT ★ 114 updated 2026-08-11
/plugin marketplace add nicknisi/claude-plugins
/plugin install essentials@nicknisiWhat squad-review does
The skill inspects git diffs across six specific areas: security, correctness, conventions, tests, architecture, and duplication. Instead of dumping raw output, it routes every spotted issue to an adversarial verifier instructed to disprove the claim. Only findings that survive this skeptical second pass reach the final review.
Execution relies on Claude Code's Workflow tool running workflow.js. Running this full cycle consumes 12 agents. Environments lacking Claude Code's Workflow capability cannot execute this script, which deliberately stops execution instead of falling back to a superficial check.
Results group by severity with blockers placed first, displaying file paths, line numbers, and actionable remediation steps. The generated report also states the number of refuted findings, items where verification timed out, and any lenses that crashed.
What it helps with
- run a thorough branch audit before merging to
main - weed out false positives by refuting AI findings before presenting them
- evaluate changes across architecture, security, and test coverage in parallel
- inspect uncommitted git diffs while highlighting relevant untracked files
Example prompts for the agent
- Run a squad-review on this branch before shipping
- Do a deep adversarial review of my diff against main
- Review current uncommitted changes with squad-review
How to install squad-review
1Plugin marketplace
/plugin marketplace add nicknisi/claude-plugins/plugin install essentials@nicknisiThis installs the whole essentials plugin with all its skills
2Skills CLI (Vercel)
npx skills add nicknisi/claude-plugins --skill squad-review -a claude-code3Manually: put the files in place
~/.claude/skills/squad-review/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/squad-review/
1Skills CLI (Vercel)
npx skills add nicknisi/claude-plugins --skill squad-review -a codex2Manually: put the files in place
.agents/skills/squad-review/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add nicknisi/claude-plugins --skill squad-review -a cursor2Manually: put the files in place
.cursor/skills/squad-review/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install nicknisi/claude-plugins plugins/essentials/skills/squad-review2Manually: put the files in place
.github/skills/squad-review/SKILL.md1Skills CLI (Vercel)
npx skills add nicknisi/claude-plugins --skill squad-review -a gemini-cli2Platform CLI
gemini skills install https://github.com/nicknisi/claude-plugins --path plugins/essentials/skills/squad-reviewManually: put the files in place
.devin/skills/squad-review/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about squad-review
How to install squad-review in Claude Code?
Register the marketplace via /plugin marketplace add nicknisi/claude-plugins, then run /plugin install essentials@nicknisi. Alternatively, execute npx skills add nicknisi/claude-plugins --skill squad-review -a claude-code.
What does squad-review do?
It evaluates branch diffs across six analytical dimensions, followed by an adversarial check that attempts to disprove every reported issue.
Why does squad-review consume 12 agents?
It launches six agents to inspect different engineering concerns and another six agents to challenge every finding.
Source text
SKILL.md · MIT
---
name: squad-review
description: >-
Review the current branch with six specialist lenses (security, correctness,
conventions, tests, architecture, duplication), then put every finding
through an adversarial verifier that tries to refute it — so what reaches you
has already survived a skeptic. Use when the user asks for a thorough,
deep, or paranoid review, a review before merging or shipping, a "real"
review, or wants to know what a review would catch that a quick pass misses.
Costs 12 agents; for a fast single-pass read, use the built-in code-review
skill instead.
argument-hint: Optional — what to review (defaults to the branch diff vs main)
---
# Squad Review
Six lenses in parallel, each finding then attacked by a verifier that assumes it
is wrong until proven otherwise. The verification is the point: an unverified
fan-out produces six sections of plausible noise, and you stop reading it by the
third review.
## Run it
Resolve the script path, then hand it to the Workflow tool:
```bash
echo ${CLAUDE_PLUGIN_ROOT}/skills/squad-review/scripts/workflow.js
```
```
Workflow({
scriptPath: "<the path printed above>",
args: { scope: "<git command>", label: "<human description>" }
})
```
## Pick the scope yourself
Read the git state and choose — do not open a menu:
- On a feature branch with commits → `git diff main...HEAD` (the normal case)
- On main, or a branch with only uncommitted work → `git diff HEAD`
- Untracked files present and relevant → name them in `label` so reviewers read
them in full; a diff won't show them
- Nothing to review → say so and stop, don't launch the workflow
Ask only when two scopes both have substantial content and you genuinely can't
tell which one the user means. If the user named a scope, use it.
## Report the result
`findings` are ranked and already verified. Present them grouped by severity, not
by lens — the lens is metadata, not structure. Lead with blockers. Give each
finding its file:line, what breaks, and the fix direction.
Three fields decide whether the review can be trusted, and all three must reach
the user:
- **`refuted`** — say how many and why, in one line. That number is the evidence
verification happened. Hiding it makes this look like every other review.
- **`unverified`** — findings whose verifier never ruled on them. Report them
under a separate heading that says exactly that. They have _not_ survived a
skeptic, so never merge them into the findings list.
- **`lenses_failed`** — non-empty means the review is incomplete. Name the lenses
that died. Never present a partial review as a clean bill of health.
Do not re-rank or soften what the verifiers confirmed, and do not write the
report to a file unless asked.
## Requirements
Needs Claude Code's Workflow tool. This skill also ships to Pi via the repo's
`pi.skills` glob, where that tool does not exist — there it will not run, and
that is the intended failure rather than a silent degradation to something
weaker.
© Nick Nisi · MIT · nicknisi/claude-plugins