Argue Both Sides Before You Decide — With an AI Skill Built to Disagree

If you're a founder, an operator, a chief-of-staff, or anyone who has to make a call that has a real downside and no obvious answer — council-of-high-intelligence is the free, open-source skill built for exactly that moment. You type /council <your question> into your AI assistant and it convenes several AI perspectives, forces them to disagree with each other, and hands you back a structured verdict.

It runs inside Claude Code, Codex, Gemini CLI, or OpenCode. The install takes two commands. And it's designed around one strong opinion: a long AI deliberation should not be used to decorate a decision you already made.

(If you don't use one of these AI assistants yet, Claude Code is the easiest way in.)

At a glance

What it actually is

A structured deliberation protocol packaged as a skill. When you invoke /council <your question>, the tool assigns your question to several deliberately-different AI personas, keeps their opening positions independent (so they don't just agree with the first one to speak), forces them to cross-examine each other, and produces a verdict.

The verdict is unusual. It leads with what remains unresolved — not with the recommendation. Then it names a recommendation, but pairs it with acceptable compromises, kill criteria (the conditions under which you should abandon the plan), and one concrete next step. When the personas remain split, the tool refuses to smooth that split into consensus — it hands the split back to you.

Three modes are available. Full runs the whole protocol — independent analysis, cross-examination, final stance, synthesis. Quick runs a faster version when broad perspective matters more than adversarial cross-examination. Duo pairs two opposing personas for questions where one tension defines the call.

What you get

• /council <question> — one command, invoked in plain English from your AI assistant's chat
• Named triads for specific domains — you can narrow to three relevant lenses instead of the whole council. Available domains include architecture, strategy, ethics, risk, shipping, product, founder, ai-product, ai-safety, systems, design, economics, and bias
• Decision field notes — before the council convenes, you write down facts, inferences, assumptions, unknowns, constraints, evidence, reversibility, and deadline. This alone kills a lot of bad questions before they waste anyone's time
• Evidence labeling in the transcript — every claim is tagged FACT, INFERENCE, or something weaker, so you can see exactly which parts of the deliberation are grounded and which are speculation
• Protocol enforcement — the tool actively checks for premature agreement, repeated claims, missing dissent, and unsupported confidence, and pushes back on the personas when they slip
• Kill criteria in every verdict — the tool refuses to hand back "do this" without also naming "and stop doing it if X happens"
• Split verdicts are preserved — when the council can't agree, you get "here are the two positions and why they don't reconcile" — not a fake middle-ground answer that will fall apart in a real meeting
• Custom councils — you can pick specific named personas by adding a short list to the command, like /council --members torvalds,ada Should we ship this? It's a two-word add

Why this matters for anyone making hard calls

Three things separate this from just asking your AI assistant "what should I do?"

First, the tool forces disagreement. Everyone who has used a chat AI on a hard decision knows the pattern — the model tends to converge toward your framing, restate your options in nicer language, and hand back a soft recommendation. That's the opposite of what you want when you're trying to argue both sides of a real problem. Council-of-high-intelligence solves this structurally — by how the deliberation is run, not by wording the request cleverly. The personas open in isolation, then cross-examine, then get pushed if they agree too fast.

Second, the field notes step is a decision-hygiene forcing function. Before the council runs, you have to write down what you actually know, what you're assuming, and what the deadline is. Half the value of the tool is that half the "hard decisions" people bring to it dissolve at this step, because the person realizes they haven't done the underlying work yet. The other half get sharper.

Third, the output shape is designed for accountability, not consensus. A verdict that leads with "here is what remains unresolved" and includes explicit kill criteria is a fundamentally different artifact from "here is what I recommend." It survives contact with a real meeting. It gives you something to point at in three months when the decision needs to be revisited.

How it compares to just asking your AI

Named alternatives, roughly in the space:

• Just typing your question into ChatGPT / Claude / Gemini — the default. Works for easy questions, fails for hard ones because of the single-perspective smoothing described above. Council is what you reach for when you catch yourself using the base chat as a rubber stamp.
• Multi-agent frameworks like LangGraph, CrewAI, AutoGen — you could build a version of this yourself with these. They're the underlying developer toolkits. Council is the pre-built, opinionated version — you install it in 2 minutes instead of building it in 2 weeks.
• Formal decision techniques (premortems, red-teaming, six thinking hats, devil's advocate) — the intellectual ancestors. Council is one attempt to encode those disciplines into an AI-executed protocol so you don't have to convene an actual meeting to get the benefit.
• claude-obsidian (9,601 stars) — different job. Obsidian brain is for organizing what you know; council is for deciding what to do with it. Complementary — use them in sequence for high-stakes calls.

Verdict

Install. No serious caveat.

For anyone who runs Claude Code, Cowork, Codex, Gemini CLI, or OpenCode, this is a two-minute install of a genuinely useful skill for the specific class of moments where you have to make a real call. MIT license means you can use it anywhere, including inside a company. There's no external service to sign up for and no additional cost — the tool uses whatever AI provider you're already using for your assistant.

The one caveat is behavioral, not technical: the tool works only if you actually use it for genuinely hard questions. Using it to legitimize decisions you've already made produces long transcripts that read like manufactured support. The tool's field-notes step and kill-criteria structure are trying to keep you honest about that; they can't fully prevent the misuse.

This post is what we learned from reading the project. A hands-on install-review — running /council on three real decisions and reporting what came back — is queued as a follow-up.

For your agent to install this

You can skip this section — it's the install details your AI agent needs. If you're using Claude Code or Cowork, say "install council-of-high-intelligence for me" and the agent will read the commands below.

System / OS

• macOS, Linux, or Windows — the skill is shell-based and shell-compatible; no runtime beyond your existing AI assistant
• Claude Code / Cowork: the plugin marketplace install below works with zero additional setup
• Codex / Gemini CLI / OpenCode: requires a working git and bash — both preinstalled on macOS and Linux, available via WSL2 or Git Bash on Windows

API keys — required for any use

None. The skill runs inside your existing AI assistant and uses whatever model provider you've already configured there. No new key, no new account.

Install commands

Claude Code / Cowork (recommended, single-line install):

/plugin marketplace add 0xNyk/council-of-high-intelligence
/plugin install council@council-of-high-intelligence

Codex CLI:

git clone https://github.com/0xNyk/council-of-high-intelligence.git
cd council-of-high-intelligence
./install.sh --codex-only

Gemini CLI:

git clone https://github.com/0xNyk/council-of-high-intelligence.git
cd council-of-high-intelligence
./install.sh --gemini-only

OpenCode:

git clone https://github.com/0xNyk/council-of-high-intelligence.git
cd council-of-high-intelligence
./install.sh --opencode-only

Restart the target client after installation. The command remains /council on every supported host.

First-run usage

/council Should we open-source our agent framework?
/council --quick Should we add caching here?
/council --duo Should we use microservices or a monolith?
/council --triad strategy Where is our defensible advantage?

The full canonical usage guide, including the routing table for named triads (architecture, strategy, ethics, risk, shipping, product, founder, ai-product, ai-safety, decision, systems, uncertainty, design, economics, bias), lives in the project's SKILL.md file at the repo root.