CAPTURE auto+gate brain-build

Your company, in one brain

The file

take it and run it.

One markdown file: the goal prompt, the verifier prompt, and the stopping condition. Read it before you run it. MIT licensed, no account, nothing to sign up for.

Download brain-build.md View the raw file

curl -fsSL https://vibemarketers.ai/loops/brain-build.md -o .claude/loops/brain-build.md

The loop

copy it.

Your agents keep producing generic work because every task starts from an empty prompt. Someone describes the company again, pastes a few links, and hopes. This loop writes that context down once, as files your agents read before they make anything.

Use it when you are setting up any agent workflow for a company and you are tired of re-explaining the business. Run it first: every other loop here reads what it produces.

You need your website, whatever docs you have, and 20 minutes to answer questions. No CRM or analytics access required, though it gets sharper if you have them.

You get seven markdown files: positioning, ICP, buyer questions, claims you can prove with their sources, competitive context, content history, and a voice spec pulled from your own best writing. They live in your repo, not on a platform.

The gate is you reading the claims file before anything downstream uses it. That is the file that decides what your agents are allowed to assert.

Goal prompt

Build the Brain for <domain>. Read the site, the docs, and any exports in
/inputs. Write Brain/positioning.md, icp.md, buyer-questions.md,
claims-and-proof.md, competitive-intel.md, content-history.md, and voice.md.
Every claim in claims-and-proof.md names a source URL and a date.

Verifier prompt

Open every file in Brain/. FAIL if any file is missing, if any claim in
claims-and-proof.md has no source URL, if voice.md has no banned-word list,
or if any file contains an em dash. Otherwise PASS.

Stopping condition

Seven Brain files exist, the verifier marks PASS, and a PR is open.

MCPs it needs

none required

Want everything at once

The installer writes the three starter skills and every published loop into .claude/. Read the script first.

curl -fsSL https://vibemarketers.ai/install.sh | bash

Or have us run it on your site →

The receipt

Receipt brain-build 2026-09-21

Baseline: No Brain existed. Brain/ held one file, voice.md, written by hand during Goal 1. Every other loop started from an empty context and guessed at positioning.
Hypothesis: A Brain built by reading both properties in full lets downstream loops cite positioning, ICP and proven claims instead of inventing them.
Intervention: Read supermarketers.ai and vibemarketers.ai in full on 2026-09-21, plus the repo's own content. Wrote positioning.md, icp.md, buyer-questions.md, claims-and-proof.md, competitive-intel.md and content-history.md, and extended voice.md with voice evidence quoted from Gen's own writing rather than invented.
Result: 7 of 7 Brain files exist. 12 provable claims recorded, each with a source URL and a date, and 5 claims explicitly listed as unprovable until a named condition is met. The verifier failed the first pass: 3 claims cited a repo path rather than a resolvable URL. Fixed, then PASS.
Commercial signal: none
Attribution limits: This loop produces an input, not a marketing outcome. No citation or traffic movement is attributable to it. The Brain is built from two websites and the repo, not from sales calls or a CRM, so the buyer-questions file is inferred from published copy rather than observed in conversation.
Learning: The claims file is where the loop earns its keep. Separating what can be proven from what cannot forced 5 claims off the site that would otherwise have read as fact, including every citation count.
Next intervention: Rebuild the Brain from sales calls and a CRM export rather than published copy, so buyer-questions reflects what buyers actually ask rather than what the site answers.

1 human touches

22 minutes

0 errors

raw run logopen →

Sources: run log, lab/runs/2026-09-21-brain-build.json, collected 2026-09-21 , window: single run ; supermarketers.ai, read in full, collected 2026-09-21 , window: homepage as published on the date read .

One entry per run, append-only, newest first. Corrections append a dated line.

Inside a company

where the lane changes.

VibeMarketers tests the loop. SuperMarketers runs it on a real team.

Problem: Every agent task starts from a different idea of what the company is, so output drifts.
Source of truth: The Brain directory: positioning, ICP, buyer questions, claims and proof, competitive intel, content history, voice.
Permitted actions: Read the website, the docs, the CRM export, and past content. Write Brain files. Open a PR.
Required approvals: One PR approval from the marketing lead before the Brain is used by any other loop.
Owner: Head of marketing
Metric: Share of downstream loop outputs that cite a Brain file instead of guessing.

On a real team the Brain reads from the CRM and the product docs, so the source of truth is the company's own records, not a website crawl.

Running this on your own site is one person reading a diff. Running it inside a company changes four things, and the guide covers them. When it stops being a file you copy and becomes a build, SuperMarketers does the same loop model inside your stack, on your keys.