Idea Council: A Free Claude Code Skill That Finds and Tests SaaS Ideas

Free, no sign-up: download the skill (zip, 29 KB) · download the prompts (Markdown). MIT licence: use it, change it, share it.

Most "SaaS ideas" lists are guesses. Someone imagines a problem, imagines a buyer, and imagines that the buyer will pay. Idea Council starts from the other end: what people already complain about in public, what they already pay for, and what they hack together because nothing fits.

It is a skill for Claude Code. You describe yourself and your market in a few lines, and a team of agents goes to work: scouts collect evidence, miners tag it, ideators propose ideas, and a council of six roles argues about them in rounds until the chair can recommend one, with the quotes, competitors and a test plan behind it.

How the Council Works

The run has five phases. Every claim in it has to point to a saved piece of evidence or a page an agent actually read. Anything else is labelled INFERRED, and made-up customer opinions never count as evidence.

Scout. Six scouts each fill a shared evidence store from one family of free, public sources: Hacker News comments, 1-3 star App Store reviews, Stack Exchange "looking for software that..." questions, GitHub issues, Reddit through search results, and money signals such as freelance gigs and review sites.
Mine. Miners read every item and tag it: the pain, who would pay, how intense it is, and signals like spend, workaround and switching. An analyst merges the tags into the 15 strongest pains, scored with spend and workarounds weighted most. A pain needs at least three independent sources to reach the top 10.
Ideas. Three ideators look at those pains through different lenses: pain first, money first, and founder fit plus what current AI makes newly possible. Each proposes six ideas with evidence ids.
Council. Six roles score every idea on their own question, then argue:

Role	The question it answers
Skeptic	Why will this fail? Counter-evidence, dead markets, weak sources.
Competitor analyst	Who already sells this, at what price, and what do their bad reviews say?
Market sizer	Bottom-up: how many reachable buyers × a realistic price? No "$5B market".
Builder	Can you ship a sellable version in 4-6 weeks with your stack?
Distribution	Where exactly do the first 10 paying customers come from, without ads?
Customer voice	Do the saved quotes show real pain, spend and switching for this buyer?

After each round the chair sums the scores, lowers them for kill arguments nobody answered, cuts the list (to at most 8, then 4, then 3), and sends up to three scouts to settle the open questions with new evidence. Arguments get settled by searches, not by the loudest agent.

Decide. The chair writes a decision memo: the recommended idea, two runners-up and why they lost, five verbatim quotes with links, competitors and prices, the first 10 customers, a 7-day validation plan with success and kill thresholds, an MVP brief, and what would change the council's mind. A script turns it into one HTML report where every quote links to its source.

The council stops when the leading idea survives a round with no new kill argument and a clear lead, or when it reaches the round limit. The report says which.

What One Real Run Looked Like

I ran it for myself on 9 October 2026 with an open search. These numbers come from that run's files:

3,018 evidence items collected, 1,156 of them App Store reviews, and 2,055 tagged by the miners.
15 pains ranked and 18 ideas proposed.
3 debate rounds: the chair cut 15 ideas to 7, then 4, then 3, and 6 evidence requests were answered between rounds.
About 70 minutes from the written brief to the finished report.

I'm not publishing the winning idea here. What matters for you is the shape of the result: one idea, the strongest argument against it, and a test you can run this week before writing any code.

Install It in Two Minutes

You need Claude Code and Node 18 or newer. The scripts use only Node's built-in modules, so there is nothing else to install. The GitHub CLI gh, logged in, adds the GitHub issues source; without it that source is skipped.

Option 1: by hand. Download the zip and unzip it into your skills folder, so you end up with ~/.claude/skills/idea-council/SKILL.md (on Windows, %USERPROFILE%\.claude\skills\idea-council\SKILL.md).

Option 2: one command. On macOS or Linux:

curl -L -o idea-council.zip https://www.elhabib.dev/downloads/idea-council-skill.zip && mkdir -p ~/.claude/skills && unzip -o idea-council.zip -d ~/.claude/skills

On Windows, in PowerShell:

Invoke-WebRequest https://www.elhabib.dev/downloads/idea-council-skill.zip -OutFile idea-council.zip; Expand-Archive idea-council.zip -DestinationPath "$HOME\.claude\skills" -Force

Restart Claude Code and ask "what skills do you have?" to check that Idea Council is listed. To check the scripts, run the built-in smoke test (no network needed):

node ~/.claude/skills/idea-council/evals/test_scripts.mjs

Before Your First Run

A full run is big: roughly 40 to 55 agents over 30 to 90 minutes, and a large share of your Claude usage limit. The skill tells you this and waits for your yes before it starts. For a first try, use the fast prompt below.

The council runs best with Claude Code's Workflow tool, which runs many agents in the background. That is why the prompts end with "Run it as a workflow". If your Claude Code doesn't have that tool, the skill runs the same phases with ordinary subagents. It is slower, and it tells you before it starts.

Prompts for Different Use Cases

Copy one, replace everything in [brackets], and paste it into Claude Code. All eleven are also in the prompts file and inside the zip.

1. I don't know what to build

Use the idea-council skill. Open search: I don't have an idea yet.
About me: [your skills, e.g. "full-stack developer, Next.js and Python"], [solo or team], based in [country],
[hours per week] a week, budget [amount] for the first 3 months.
I want B2B, [price range, e.g. 30-150 USD a month], and [revenue goal] a month within a year.
Deal-breakers: [e.g. no marketplaces, no hardware, no regulated health data].
Run it as a workflow.

2. Find an idea inside one industry

Use the idea-council skill for one market only: [industry, e.g. independent dental clinics / tour operators /
property managers / restaurants]. Scouts should stay inside that market: its apps' 1-3 star reviews, its forums and
subreddits, its job posts and freelancer gigs.
About me: [skills, location, budget, solo or team]. I already know this market because [your connection to it].
Run it as a workflow.

3. Try to kill my own idea

Your idea enters the debate as one candidate among the ones the council finds. It can lose, and that is the point.

Use the idea-council skill to test an idea I already like: "[your idea in one sentence]".
Put it in the debate as a candidate next to the ideas the council finds itself. The skeptic and the competitor
analyst should try hard to kill it. If it loses, tell me plainly which idea beat it and why.
About me: [skills, location, budget].
Run it as a workflow.

4. Pick between ideas I already have

Use the idea-council skill to choose between my ideas:
1. [idea A]
2. [idea B]
3. [idea C]
Scout evidence for the pains behind all three, add them as candidates, and let the council rank them against
whatever else it finds. I want one winner, the strongest argument against it, and a 7-day test.
About me: [skills, location, budget].
Run it as a workflow.

5. A local market or another language

The App Store scout can read another country's store, which matters when your buyers don't write in English.

Use the idea-council skill for small businesses in [country / city]. Collect App Store reviews from the [two-letter
country code, e.g. fr, ma, de] store, search in [language(s)], and look at local public forums and job boards there.
Pricing must fit local budgets.
About me: [skills, budget], I sell in [language].
Run it as a workflow.

6. A fast, cheaper first run

Use the idea-council skill with a small run: scoutTarget 120 and maxRounds 2.
About me: [one or two lines]. Market: [open, or one industry].
Show me the estimate first. Run it as a workflow.

7. Old pains that AI makes solvable now

Use the idea-council skill and weight the founder-fit / AI-leverage lens: I want pains that were too manual or too
expensive to automate before and are now solvable with current AI models (documents, calls, email, images,
spreadsheets). The skeptic should check that the AI part works today, not in a demo.
About me: [skills, which AI APIs you use, budget].
Run it as a workflow.

8. Turn my service into a product

Use the idea-council skill. I run a [service, e.g. web design / bookkeeping / video editing] business for
[client type]. Look for the recurring manual work my clients and people like them pay for, and find a SaaS I can
sell to my existing clients first.
About me: [team size, current clients, skills].
Run it as a workflow.

9. Resume a run that stopped

In the same Claude Code session, finished agents are reused:

The idea-council run stopped before the end. Resume the same workflow from its run id so finished agents are
reused, then save council.json and render the report.

In a new session, the evidence already collected stays on disk and duplicates are skipped:

Re-run the idea-council skill with the run folder [run folder path] and the same brief. Check what is already in
the evidence store first and only scout what is missing.

10. After the council: test it this week

Read the Idea Council report in [run folder path]. Turn its 7-day validation plan into: a one-page landing page
(headline, subhead, three benefits, one call to action, a waitlist or "book a call" button), the interview script,
and the list of exact places to post with a first message for each. Keep every claim to what the evidence supports.

11. After the council: plan the build

Read the Idea Council report in [run folder path]. Turn the MVP brief into a build plan for [your stack]: data
model, pages, integrations, and a week-by-week plan for [number] weeks. Put the riskiest technical part in week 1.
Don't start building until the 7-day test passes its success threshold.

What It Will Not Do

It doesn't prove demand. It ranks ideas against public evidence. Complaints are not purchases. The 7-day test with real people is what decides, and the report says so.
It doesn't get around blocks. It never logs into anyone's account and never bypasses a CAPTCHA. Reddit blocks direct reading, so the Reddit scout only uses search results. Sites that block reading are skipped and listed in the report.
It doesn't replace knowing your market. The brief is a filter: the more honestly you describe your skills, budget and deal-breakers, the more useful the winner is for you rather than for someone else.

What's in the Zip

SKILL.md: the instructions Claude Code follows, with the checklist for a run.
workflows/council.js: every role, prompt, output schema and stop rule. Change the roles here if you want a different council.
scripts/collect.mjs, scripts/evidence.mjs, scripts/report.mjs: the free collectors, the evidence store, and the HTML report.
references/roles.md: the scoring rubric, stop rules, source limits and how to tune a run.
PROMPTS.md, README.md, LICENSE (MIT), and evals/test_scripts.mjs.

If you run it, I'd like to hear what it found and where it got things wrong. Send me a message on LinkedIn.