multi-agent-review Skill: What It Does and How to Install It

Reviews a spec or plan before you build it: six reviewer agents, a juror on disagreements, and a verdict that gates the next step.

Installs: 20.5K; on skills.sh
Stars: 82; on GitHub
Files: 5; besides SKILL.md
Updated: Aug 2026; last repo commit

The SKILL.md has the agent run two model tiers over three topics: completeness, alignment and risk. If the tiers disagree on a finding, one higher-tier juror decides. The result sorts findings into Blockers, Warnings and Observations, and blockers stop the next step until you fix them or log an override.

Run /multi-agent-review spec before writing a plan, or /multi-agent-review plan before executing it. It reviews specs and plans, not code. By default it looks for the newest file in docs/superpowers/specs/ or plans/, so outside that layout you pass a file path.

Good for

Developers who write specs and plans before coding
Teams using the Superpowers planning workflow
Anyone who wants a second opinion on a plan

Install multi-agent-review

Claude Code

Copy this prompt.

Install the multi-agent-review skill in Claude Code. Run this command:

/plugin marketplace add riekelt/multi-agent-review
/plugin install multi-agent-review@multi-agent-review

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Pick one route. The plugin manifest sets the model tiers (haiku, sonnet, opus on Claude Code). The skills CLI route installs only the skill folder, and SKILL.md says it falls back to haiku, sonnet and opus when it finds no manifest. The plugin has no hooks or MCP config in the repo tree. Without the Superpowers skills, pass an explicit file path (see the FAQ).

From the project README. Source.

What's in the skill

When it loads

Use when a spec or plan needs review before execution begins - "review my spec", "is this plan ready to execute", "check this before I build it", "second opinion on this plan" - and specifically before writing-plans (spec mode) or before subagent-driven-development (plan mode). Reviews specs and plans, not code. Panels six reviewers across two model tiers and three topics, invokes a reasoning-tier juror only when the tiers disagree, and gates the next workflow step with a fail-closed Blockers / Warnings / Observations verdict.

Sections in SKILL.md

When to invoke
Invocation syntax
Coordinator steps
Step 1: Locate the artifact
Step 2: Validate, then read the artifact
Step 3: Resolve model tiers, read companion files, read project rules
Step 4: Build the six agent prompts
Step 5: Dispatch six agents in parallel
Step 6: Compare pairs, identify contested findings
Step 7: Invoke juror (ONLY if CONTESTED_LIST is non-empty)
Step 8: Quorum check + compile final verdict
Step 9: Decision gate
Panel failure semantics
Model assignments
Scope limits
Companion files

Key facts

Author: riekelt (repo owner; plugin.json names riekelt as author)
License: MIT (README and plugin.json; the repo has no LICENSE file)
Installs on skills.sh: 20,487 (Nov 2, 2026)
GitHub stars: 82 for the whole riekelt/multi-agent-review repo (Nov 2, 2026)
Last repo commit: Aug 22, 2026
What it includes: 6 files including SKILL.md: 4 reviewer prompt files and 1 project-rules example
Needs: Model access for three tiers (haiku, sonnet and opus on Claude Code, set in plugin.json); the default file search expects the Superpowers folder layout
Works in: Claude Code, Codex and Cursor

What it includes

Review a spec after brainstorming, before writing the plan.
Looks for the newest file in docs/superpowers/specs/ unless you pass a path. On a clean result it starts writing-plans.
Try: “/multi-agent-review spec”
Read · /multi-agent-review spec
Review a plan after writing it, before execution.
Looks for the newest plan in docs/superpowers/plans/ and loads the linked spec as context. On a clean result it starts subagent-driven-development.
Try: “/multi-agent-review plan docs/plans/billing.md”
Read · /multi-agent-review plan
A fast and a standard tier each review completeness, alignment and risk.
The six agents run in parallel on the full artifact, using the prompt files in the skill's agents folder.
Read · Six-agent panel
A reasoning-tier model rules on findings the two tiers disagree on.
Called only when there is a disagreement. If the juror fails, every contested finding becomes a blocker.
Read · Juror
Run the fast tier only: three agents, no juror.
The skill refuses --fast and runs the full panel when the artifact mentions authentication, security, secrets, payment, billing, migration, data deletion or production infrastructure.
Read · --fast
Sort findings into Blockers, Warnings and Observations.
Blockers stop the next step; warnings ask you to fix or accept. It also prints a JSON block with the verdict and panel health.
Read · Verdict gate
Proceed past blockers and record a git note.
Appends a note to the artifact's latest commit. Re-runs are capped at 3 iterations.
Write · Override log
Give the alignment and risk agents your standing rules.
Copy assets/project-rules.example.md to your project root. Without it the skill falls back to CLAUDE.md or AGENTS.md.
Read · project-rules.md

From the official docs

Permissions

Medium risk

The skill sends your whole spec or plan to several model calls, and an override writes a git note to your latest commit. Keep secrets out of the documents you review.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

It sends your full spec or plan to up to seven model calls
The SKILL.md reads the whole artifact and passes it verbatim to six reviewer agents, plus a juror when the two model tiers disagree. Each re-run repeats this, up to 3 iterations. Use --fast for low-risk documents; the skill refuses --fast for security, billing, migration or production topics. github.com
An override writes a git note
If you override blockers, the SKILL.md appends a git note to the artifact's latest commit with the timestamp, count and blocker titles. Nothing else in your repo is edited. github.com
A clean review starts the next skill on its own
On a clean verdict the SKILL.md auto-invokes writing-plans (spec mode) or subagent-driven-development (plan mode). Both are Superpowers skills, not part of this plugin. github.com
It reads your project rules
It reads project-rules.md, or CLAUDE.md or AGENTS.md as a fallback, and gives them to the reviewers. github.com

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Turning on read-only	The SKILL.md says it surfaces findings only and does not repair artifacts.
Scopes	The spec or plan file you point it at project-rules.md, CLAUDE.md or AGENTS.md in your project Git notes on your latest commit (on an override)

Questions

What does multi-agent-review do?

It reviews a spec or plan before you execute it. Six reviewer agents on two model tiers check completeness, alignment and risk, and a higher-tier juror settles disagreements. It does not review code.

How do I install multi-agent-review?

Use the install box above. On Claude Code the plugin route is the one the README leads with, and it carries the model tier settings. Pick one route, not both.

Does it need the Superpowers skills?

It is built to slot into the Superpowers workflow and auto-invokes two Superpowers skills by name, writing-plans and subagent-driven-development. The README says that without Superpowers you invoke it directly and pass an explicit file path, because the default search assumes the Superpowers folder layout.

Does it work in Codex and Cursor?

Yes. The README has a copy-the-plugin install for each and says both read their model tiers from their own plugin.json. It documents no Claude app install, and the skill runs local commands, so the install box has no Claude app tab.

What does a review cost?

A full review makes six reviewer calls, plus one juror call when the tiers disagree. Each re-run repeats that, up to 3 iterations. The README says --fast uses the fast tier only, with three agents and no juror.

Is multi-agent-review safe to install?

It reads your spec or plan and sends it to model calls, and on an override it writes a git note to your latest commit. The SKILL.md says it does not repair artifacts. Our advice: keep secrets out of the documents you review, since the whole text goes to every reviewer.

Information

updated Nov 2, 2026

Repository: riekelt/multi-agent-review
Folder: plugins/multi-agent-review/skills/multi-agent-review

riekelt/multi-agent-review skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 4

multi-agent-review README · github.com
multi-agent-review plugin.json · github.com
marketplace.json · github.com
multi-agent-review SKILL.md · github.com