Claude Council Khalifa skill
A Claude Code skill that tests a business idea with a four-agent council and rules: build, fix first or kill.
- 116 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The skill only reasons and writes notes into the council folder in the working directory.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/claude-council-khalifaDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill only reasons and writes notes into the council folder in the working directory.
This is third-party code. Review the repository files before installing.
What it does
The skill answers whether an idea is worth building before months and money are spent on it. An ordinary model query gives a balanced pros and cons list that flatters the author. Here the idea is first rewritten into a flat paragraph with no persuasion, then four agents, each in its own fresh context, do their job. The Believer builds the strongest honest case for the idea: who is in pain, what it costs today, why now. The Skeptic tries to kill it: who will not pay, real competitors including a spreadsheet, and the assumption the author cannot see. The Investor counts money: unit economics, the first ten customers and the cheapest demand test this week. The Judge makes them fight, rules BUILD, FIX FIRST or KILL and adds one ten-minute action. Rulings accumulate in council/council.md, and on the next call the Judge checks whether the verdict moved.
Who it is for. Early-stage founders and entrepreneurs who need a hard check of an idea before development or launch.
Good fit when
- There is a subscription service idea and you must see whether the first ten customers would pay.
- The author is too attached to the project and expects an honest refusal rather than support.
- After a demand test you return to check whether the verdict changed.
- The idea is written in Arabic and the council answers in the same language.
Not a fit when
- You need a finished financial model or market sizing: the Investor tags numbers as assumptions.
- The idea is already in development and the decision is made: the council will neither support nor cancel it.
- You need technical implementation advice: the council assesses demand.
Example request
claude council: an app that sends subscribers ready recipes from their fridge contents, is it worth building in three months?Limitations
The verdict depends on the idea text and the model: the skill does not verify figures and tags Investor numbers as assumptions or with a source link. The agents have no market access, so you must run the demand check yourself as the ten-minute action suggests. Memory lives in the council folder next to the project and is written to the current directory. If the Judge says BUILD every time, the author advises treating the prompts as broken. The MIT license is in the repository, and the README is in English and Arabic.
How to disable. Delete the claude-council folder from ~/.claude/skills. Remove the council folder with the ruling history in your working directory separately.
FAQ
How do I install the skill?
Clone the repository, then run `mkdir -p ~/.claude/skills` and `cp -r claude-council-khalifa/skills/claude-council ~/.claude/skills/`. Then open a new Claude Code session.
How do I call the council?
Write `claude council:` followed by the idea in one paragraph. In Arabic the `المجلس:` form works.
Which verdicts exist?
BUILD, FIX FIRST and KILL. The Judge adds one action you can do in ten minutes today.
Where are results stored?
The full transcript of each run is in `council/sessions/`, and short memory is in `council/council.md` in the working folder.
How do I return to an idea after testing?
Write `claude council: what changed?` and describe test results. The Judge rereads history and says whether the verdict moved.
Discussion
No questions or reviews yet.
Related
SkillsThe /court skill puts an idea on trial: a prosecutor, a defense, 12 jurors and a judge instead of Claude’s usual agreement.
Eleven skills that test a business idea before launch: competitors, a panel of 100 buyers, pricing, a financial model and a 30-day plan
Idea Validation Agents: app idea validation
idea-validation-agents
Skills for validating an app idea: trends, competitors, pricing, retention, acquisition cost and a final verdict scored 0 to 100
A skill that scores a startup idea on 14 dimensions with web research: scorecard, assumption ledger and HTML report.
Ten Claude Code commands based on The Minimalist Entrepreneur: from finding a community and validating an idea to pricing, marketing and company values
A plugin of skills for startup validation and planning: market research, competitor analysis, positioning and pitch
README in short
The repository has a single skill and a README in English and Arabic. The skill describes four roles, built-in guardrails against flattery and automatic reruns of weak answers. Run results are written to the council folder.
README badge
Original SKILL.md
---
name: claude-council
description: Stress-test a business idea with a 4-agent council — the Believer argues FOR it, the Skeptic tries to kill it, the Investor asks if real money shows up, and the Judge rules BUILD / FIX FIRST / KILL with one 10-minute de-risk action. Remembers every verdict in council.md so ideas can be re-judged later. Use when the user says "council", "claude council", "run the council", "is my business idea good", "should I build this", "judge my idea", "المجلس", "مجلس كلود", or pastes a business idea and wants an honest verdict. Also use for "what changed since the last verdict" / re-judging an idea already in council.md.
---
# The Claude Council — 4 agents, 1 verdict
A single model asked "is my idea good?" wants to be helpful, and helpful sounds like a balanced list of pros and cons that flatters the founder into building the wrong thing. The council fixes that by giving each agent ONE job, running each in a fresh context, and forcing a judge to rule.
Flow: flat idea paragraph → Believer → Skeptic → Investor → Judge → verdict written to `council.md`.
## Step 0 — Set up
1. Get today's date with `date +%Y-%m-%d`.
2. Memory lives in `council/` inside the current working directory:
- `council/council.md` — the shared memory (one short entry per agent per run).
- `council/sessions/<date>-<idea-slug>.md` — full transcript of each run.
Create the folder/files if missing. Read `council/council.md` if it exists.
3. Detect mode:
- **New run** — the user gave an idea. Continue to Step 1.
- **Re-judge** — the user asks what changed / wants a re-ruling on an idea already in `council.md` → go to "Re-judge mode" below.
4. Language: agents answer in the language the user wrote the idea in (Arabic idea → Arabic answers, Egyptian dialect if the user writes in Egyptian). Keep the verdict labels (BUILD / FIX FIRST / KILL) in English so the log stays searchable.
## Step 1 — Flatten the idea
Rewrite the user's idea as ONE flat paragraph, the way a stranger would describe it: who it's for, what it does, how it makes money. Strip all persuasion ("revolutionary", "huge market", "everyone needs this"). A leading paragraph is the #1 reason all four agents agree.
If the idea is missing something essential (who pays, or what it actually does), ask the user ONE question, then proceed. Otherwise don't ask — just go.
Give it a short IDEA NAME (2–5 words) for the log.
If `council.md` already has entries for this same idea, include them (just the entries for this idea) as `PRIOR COUNCIL HISTORY` in every agent prompt below, so the council remembers.
## Step 2 — Run the agents
Run each agent as a **separate sub-agent** (Agent tool, `general-purpose`) — never answer as the agents yourself in this conversation. Shared context contaminates them and they start sounding identical. Run them **in sequence**, because each one reads the ones before it. Paste the prompts exactly, filling the brackets. Tell every sub-agent: "Return only your answer. Do not write any files."
The Skeptic and Investor may use WebSearch to name real competitors and real prices. Any number they use must be tagged `(assumption)` or `(sourced: <link>)`.
### Agent 1: the Believer
```
You are the Believer. Your only job is to make the strongest honest
case FOR this idea. Not a balanced view. The best possible version.