A Claude Code skill that turns one short prompt into a multi-agent brainstorm, and a worked example: finding Jev use cases that have not been widely discussed.
The repository holds the skill, the prompt it uses, and a complete record of one real use. The record shows what the skill does and what the model decides on its own. It lists every agent’s prompt, every candidate idea, and the two ideas that were chosen.
| Page | What it shows |
|---|---|
| Brainstorm results | All 35 candidate ideas from two runs, the two chosen ideas with their tests and costs, and the ranked rejected lists. |
| Development journey | How the 14 subagents worked: stages, roles, exact prompts, who talked to whom (nobody did), and how the agent count was chosen. |
GitHub strips embedded frames from a README, so the pages open as links. Both are single HTML files in dark mode. The source files are brainstorm-results.html and DEVELOPMENT-JOURNEY.html in this repository.
The skill comes from Ethan Mollick’s essay The Dot and the Swarm (One Useful Thing, 1 Oct 2026). In the essay, a short prompt with no team description made a model start 3 agents and organize the work itself. The skill reproduces that idea in Claude Code, where the model decides how many subagents to start and what each one does.
This project is not affiliated with Ethan Mollick or with TypeSafe AI.
The skill sends the prompt below. It is adapted from the prompt quoted in Mollick’s essay, which named a single topic (“the next OneUsefulThing post”). It has two changes from that original:
Goal:
Generate candidate ideas for {TOPIC} and choose one.
Search strategy:
Explore the idea space from as many different angles as possible.
Evaluation:
Assess candidates from:
1. factual / evidentiary perspective
2. reader perspective
3. what other publications are covering
You may delegate. Organize the work as you think is best.
| Change | Reason |
|---|---|
the next OneUsefulThing post became {TOPIC} |
The skill takes any topic from the user. |
The last line was added: You may delegate. Organize the work as you think is best. |
The essay’s run used a mode that delegates by itself. Claude Code has no such mode, so the permission is stated in words. |
The skill file also adds one rule: do not define teams, roles, or agent counts. The prompt gives the goal, the search strategy, and the evaluation views. The model decides everything else.
SKILL.md to ~/.claude/skills/swarm-brainstorm/SKILL.md./swarm-brainstorm <your topic>.Jev is the first “System One” model from TypeSafe AI. It takes unstructured state and a typed question, and returns a typed answer (a choice among up to 255 options, a score, or a boolean) with probabilities. It never returns free text. The question for the brainstorm was: which Jev use cases are not widely discussed?
The same prompt ran twice with {TOPIC} set to that question.
| Run | Evidence base | Subagents | Candidates | Chosen idea |
|---|---|---|---|---|
| 1 | Open web research | 7 (3 + 4) | 20 | All-pairs contradiction scan of a policy corpus |
| 2 | Only the TypeSafe launch blog and the awesome-jev list | 7 (4 + 3) | 15 | Dead-letter-queue triage with a replay contract |
Run 1 winner. Split a policy corpus into clauses. Ask Jev about each candidate pair in both orders. Flag a pair only if both orders say “contradict” with probability 0.7 or higher. Code checks numbers and dates.
Run 2 winner. A small consumer reads each message from a dead-letter queue and asks Jev to choose: replay, replay after a fix, quarantine, or page a human. Code lets only idempotent messages replay automatically. Jev can never discard a message. The cost estimate is $44.10 for one million messages.
Nothing was built or tested. The thresholds in the results are proposals from the agents.
| File | Purpose |
|---|---|
SKILL.md |
The skill, ready to install. |
brainstorm-results.html |
Results of the Jev brainstorm. |
DEVELOPMENT-JOURNEY.html |
The journey document with all 14 prompts. |