NewAI Operating System for contractors. Get a free analysis of where AI fits your business. →
Joey Farbstein
01 / 02Skills · 6 min read · Updated Sep 28, 2026
🥚

The /hatch Skill

Turn a fuzzy idea, a half-baked plan, or a risky decision into a clear, defended result.

Install this skill · Claude Code

$claude plugin marketplace add ThetaOneMarketing/hatch && claude plugin install hatch@hatch

Then type /hatch:hatch <your problem> in Claude Code.

Or, Hatch Lite · ChatGPT, Claude, any chat AI

📄hatch-for-chat.md — paste it as the first message in a new chat

Copy the file, open a new chat, paste, type your problem. Download it instead if you prefer.

What it does

hatch takes something you can't quite pin down yet, an idea, a plan, a decision, and runs it through six stages until it comes out the other side as a thing you can defend. It interviews you until the question is locked. It brainstorms from angles you wouldn't pick and skips the three answers anyone would give. Five advisors who disagree on purpose argue it out and one chair picks. A cold auditor finds the assumption that kills it. Then it builds, and a stranger who never saw the conversation judges the result: SHIP or NOT YET, and the one biggest gap.

Ask an AI once and you get the first safe answer anyone would give in thirty seconds. Hatch is what I built because that answer kept agreeing with me, and I was still blind.

When to reach for it

You invoke it by typing /hatch:hatch (plain /hatch if you installed it any other way); the agent won't reach for it on its own. Reach for it when the obvious answer would be expensive if it's wrong, and when two smart people would give you different answers.

  • Building an idea from nothing. A business, a side project, a course, a channel. Vague in; a brief, a plan, and a first version out.
  • Seeing angles you'd never think of. Naming, positioning, pricing, hooks. Thirty angles, sorted, the weird-but-viable one starred.
  • Deciding between options. Job offers, tools, vendors, hire vs. automate, "should I launch this?"
  • Stress-testing what you already believe. A plan, a strategy, a pitch, someone else's proposal.
  • Getting a real verdict on finished work. A draft, a design, a page, code, a résumé. Judged by a stranger against what you asked for.

Skip it for lookups, syntax, and bugs with a known cause. It will tell you so and answer directly instead. Each stage also runs alone: "just brainstorm this," "just red-team my plan," "judge these three."

Two versions

Hatch for Claude Code is the full method and the one I use. Hatch Lite is the same six stages in one file you paste into ChatGPT or Claude. Lite is a real upgrade over asking once. The full version is better, and here is exactly why:

  1. Five separate brainstormers that can't see each other's ideas. One chat can't un-see its own list.
  2. Reviewers who don't know who wrote what.
  3. A critic that has never seen the conversation. Fresh context, every round.
  4. A run file on disk, so a long session never loses its place.
  5. Twenty-plus AI calls per run instead of one. More minds, wider spread, better pick.

If you don't use Claude Code, start with Lite. If you do, don't bother with Lite.

The honest part

Lite can't fully keep two of the rules. In one chat window the brainstormers can see each other and the reviewers aren't truly blind. It says so, and it uses a fresh-chat trick at the end so a genuine stranger judges the result. On small asks (a name, a choice) the gap barely shows. On big ones (a plan, a strategy) it does.

Hatch stands on published work by others: Matt Pocock's interview-in-rounds, Udit Akhouri's isolated generators, Andrej Karpathy's LLM council, Nate B. Jones's assumption audit, Matt Shumer's gauntlet loop. Credits are in the repo.

It's working if

  • You disagree with an advisor. A run with no pushback is a run you didn't need.
  • The pick isn't the first idea you had. If it is, the obvious answer was right and you know it now.
  • The risk sheet names something you hadn't written down, and you can check it before spending a week.
  • You could defend the result to someone who wasn't there.

Common questions

How long does a run take?
Quick mode: a few minutes and one round of questions. A standard run in Claude Code: twenty or so AI calls, five to ten times a single answer. Use quick for names and small choices; save standard for anything worth a day.
It asked me things I'd already said.
Scope was too big. Ask it to break the problem into pieces and hatch one. Long sessions drift.
Can I just get the verdict on options I already have?
Yes. "Judge A vs B vs C" runs only the advisor stage.
Does it work on non-business things?
Yes. Career moves, relocation, how to spend a year. It doesn't care what the subject is, only that the answer isn't obvious.
Is the ChatGPT version really free?
Yes. MIT license, no email, no signup. Same for the Claude Code version.

Where it fits

Hatch is the first skill in the Think group, and the one I reach for before anything gets built. The skills that come after it (the ones that build, sell, and run) all start from a brief Hatch produced. Its last stage, the stranger who judges the result, is also its own skill: raise-the-bar.

Free · one at a time

Hatch finds the weak idea. It does not run the business.

The tools that do are coming next. DM me "hatch" on Instagram and I'll send each one as it lands.

Previous skill
None yet
You are here
The /hatch Skill