Council skill by Warp: several models, one decision memo

Makes your agent ask several subagents on different models the same question, then merge their answers into one decision memo. Built for Warp.

Installs: 27.9K; on skills.sh
Stars: 644; on GitHub
Updated: Oct 2026; last repo commit

council makes your agent launch several subagents on different models, have each answer the same question on its own, and then merge their reports into one decision memo. Use it for judgment calls like architecture tradeoffs, risky bug fixes and "is this worth doing?" questions. It has no scripts and needs no API keys of its own. The mixed-model part only happens in Warp, which is the one app the author documents; in Claude Code, Codex or OpenCode every member comes from one model provider, and with a single model it falls back to different angles.

Good for

Developers weighing two technical approaches before committing
Reviewers who want a risky fix or PR red-teamed by more than one model
Warp users who want Claude, GPT and open-source models to disagree in the open
Teams making rollout or incident calls where people disagree

Install council

Claude Code

Copy this prompt.

Install the council skill in Claude Code. Run this command:

npx skills add https://github.com/warpdotdev/common-skills --skill council

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.

Requires Node.js.

The mixed-model council runs only in Warp. Here the skill falls back to one model with different angles, which the author does not document. Claude Code subagents run Claude models only.

From its listing on skills.sh. Source.

What's in the skill

When it loads

Run a model-diverse subagent council to investigate the same problem from multiple perspectives, compare findings, and produce a final recommendation. Use this skill whenever the user asks for a council, second opinions, multiple agents/models to evaluate one question, parallel investigation, red-team/blue-team comparison, or help deciding between competing technical approaches.

Sections in SKILL.md

Workflow
1. Frame the council question
2. Choose council members
3. Brief before launching
4. Ask for structured reports
5. Collect reports
6. Synthesize the recommendation
Final answer template
Recommendation
Why
Tradeoffs and risks
Final call
Practical notes

Key facts

Author: Warp (written by Zach Bai)
License: MIT
Installs on skills.sh: 27,926 (Nov 2, 2026)
GitHub stars: 644 for the whole repo of 29 skills (Nov 2, 2026)
What installs: One text file, no scripts
Needs: Subagent support in your client; for mixed models, Warp with a usage balance
Last commit: May 20, 2026 to this skill; repo updated Oct 28, 2026
Works in: Warp (documented); loads in Claude Code, Codex, OpenCode with one model provider only

What it includes

States the decision in one sentence: options, artifact, read-only or not, criteria.
Before launching anything, the agent pins down what is being decided and how the winner will be judged.
Try: “Run a council on keeping our REST client vs switching to the generated one.”
Read · Frame the question
Picks 3 members, each a different model and angle (Claude, GPT, open-source).
One strong Claude model checks correctness, one GPT or Codex model covers implementation and tests, and one open-source model argues against. With only one model available, it says so and varies the angles instead.
Try: “Get second opinions from multiple models on this design.”
Network · Choose members
Tells you who it will launch and why, then waits for your OK.
Every member gets the same repo path, branch, question, constraints and report format.
Try: “Council this, read-only, no commits.”
Brief before launching
Each member returns evidence, risks, cost, recommendation and confidence.
Members fill the same fixed fields, so their reports can be compared side by side.
Files · Structured reports
Reads each report and asks the same member follow-ups instead of relaunching.
Weak claims get a focused follow-up question to the member that made them.
Collect reports
Weighs reports by evidence, not votes, and writes one decision memo.
The memo leads with the recommendation, names agreement and disagreement, and splits what to do now from what can wait.
Synthesize
Recommendation, Why, Tradeoffs and risks, Final call.
The answer comes back in a short fixed layout, with confidence and unknowns called out.
Final memo

From the official docs

Permissions

Low risk

council is a plain-text skill with no scripts, so the main risk is what its subagents do: prototype councils can edit code, and your repo context goes to several model providers. Risk is low.

What it can do

Read your data
Create and change things
Delete things
Read-only modeHow to turn it on

Watch out for

Snyk warns about outside PR and issue content
Snyk (W011, risk 0.30, Sep 14, 2026) says inspecting PRs or issues written by outsiders can expose agents to prompt injection. Do not point a council at untrusted PRs without read-only mode. skills.sh
Your code goes to several model providers
The default roster sends the same repo context to a Claude model, a GPT model and an open-source model host. Check that your team allows all three. github.com
Prototype councils edit code
For implementation councils each member gets its own git worktree and branch and may make changes. Only read-only councils are told not to edit. github.com
No scripts; instructions only
The skill is one SKILL.md with no bundled scripts or packages. What it can do depends on the subagents your client launches and their permissions. github.com
skills.sh audits: Safe and Pass
Gen Agent Trust Hub rated it Safe and Socket passed it, both on Sep 14, 2026. skills.sh
Members wait for your approval
The skill tells the agent to describe the planned members and wait for your OK. Warp also shows an approval card before launching child agents by default. docs.warp.dev

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Turning on read-only	Ask for a read-only council: the skill keeps all members in one checkout and tells them not to modify files, commit, create branches or open PRs. In Warp, the run_agents permission (always_ask / never_allow) also gates child launches.
Talks to	Model providers your agent routes to (in Warp: Anthropic, OpenAI, an open-source host such as Fireworks)

Examples

Pick between two technical approaches in a codebase
Run a council on whether we should keep the current caching layer or move to the alternative in this branch. Read-only.
Red-team a risky fix before merge
Get second opinions from multiple models on this PR's fix. Red-team it for edge cases.
Decide a rollout when people disagree
Council this: ship the migration this week or wait for the next release? Criteria are rollout safety and rollback cost.
Build competing prototypes, then compare
Run an implementation council: each member builds a candidate for the retry logic in its own worktree, then recommend one.

Questions

How do I install council?

Use the install box above. The author documents the Warp route, so that method comes first. The same skill also installs in Claude Code and Codex, but without Warp it cannot mix models.

How do I use the council skill?

Ask for a council or for second opinions on one question, for example "Run a council on whether we should move this service to Postgres." The agent proposes the members and their angles, launches them, and returns one memo with a recommendation. The skill's description lists council, second opinions, red-team/blue-team and choosing between technical approaches as the requests that start it.

Is there one official "council" skill?

There is no single "council" skill, and none is from Anthropic. This page covers Warp's, which runs subagents on different models. affaan-m's ecc council uses four Claude voices for go/no-go calls. coreyhaines31's marketing-council has one model play 12 named marketers. The llm-council skills follow Karpathy's idea: rohitg00's llm-council calls real model APIs with your keys, and aiwithremy's llm-council uses five Claude advisors with different lenses. Karpathy's original llm-council is a local web app, not a skill.

Does council work in Claude Code?

It installs and loads, but the author documents only Warp. The skill launches members with Warp's child-agent tool, and Claude Code subagents run Claude models only. You get several Claude subagents with different angles, which the skill treats as its fallback and tells the agent to say so.

Does council work in Codex or OpenCode?

Not as designed. Codex runs subagents and can give them other OpenAI models, but they inherit the parent's provider, so there are no Claude or open-source members. OpenCode subagents reuse the session model, so it falls back to one model with different angles. Warp can launch Codex or Claude Code as child agents, but only when Warp is the parent.

Does council work in Claude.ai or the Claude app?

Nobody documents it. The skill works by launching subagents, and Anthropic's docs for the Claude app do not describe a way for a skill to do that.

Do I need API keys for GPT or open-source models?

Not in Warp. Warp Agent routes to Anthropic, OpenAI, Google and Fireworks-hosted models on your Warp account and usage balance. The docs do not name a plan requirement.

Is council safe?

Yes, with one caveat. It is a single instruction file with no scripts. Gen Agent Trust Hub and Socket passed it, and Snyk gave a medium warning that reading outside PRs or issues can feed prompt injection to the agents. Ask for a read-only council if you do not want members editing code.

council vs llm-council: what is the difference?

llm-council skills copy Karpathy's three steps: models answer, rank each other anonymously, and a chairman writes the verdict. Warp's council is aimed at codebase decisions. Members inspect files, return structured evidence reports, and the lead writes a memo that weighs evidence, not votes.

Can the council members critique each other?

Not by default. Round one stays independent so members do not anchor on each other. The same repo's cross-critique skill adds a second round where each member reviews the others' proposals.

Information

updated Nov 2, 2026

Repository: warpdotdev/common-skills
Folder: .agents/skills/council

warpdotdev/common-skills skills.sh

Sources

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

Show all 20

warpdotdev/common-skills README · github.com
council on skills.sh · skills.sh
Codex subagents · learn.chatgpt.com
council SKILL.md · github.com
Warp Agent harness · docs.warp.dev
Claude Code subagents · code.claude.com
davesnx/setup: adapted council SKILL.md · github.com
warpdotdev/common-skills · github.com
Add /council (#15) · github.com
Snyk audit: council · skills.sh
skills.sh search: council · skills.sh
cross-critique SKILL.md · github.com
affaan-m/ecc council · github.com
marketing-council · github.com
rohitg00 llm-council · github.com
aiwithremy llm-council · github.com
karpathy/llm-council · github.com
Warp orchestration · docs.warp.dev
Socket audit: council · skills.sh
Gen Agent Trust Hub audit: council · skills.sh