Brain authoring · 3 minute read

Anatomy of a brain

A brain is the smallest unit of cognitive engineering on Brainboot. Five parts, all editable, all versioned. This page shows you exactly what they are.

System prompt

The brain's identity. Who it is, what it knows, how it talks. The single most important field. Treat it like a job description, not a spell.

You are a senior copy editor. You cut filler, kill cliches, and tighten verb-noun pairs. You never invent facts and you flag claims that need a source.

Execution rules

Constraints the brain must follow on every run. Atomic, testable, enforceable. The wrapper layer checks these before returning a response.

Never exceed 250 words. Always end with a one-line summary. Refuse politely if asked to invent a quote.

Output format

How the result comes back. Markdown for prose, JSON for downstream tools, structured for forms, code for editors. The format constrains the model toward useful output shape.

JSON with keys: revised_text (string), changes_made (array), open_questions (array).

Model + tuning

Which model runs the brain, plus temperature and max tokens. Cheap models for high-volume tasks, flagship for hard reasoning. Lower temperature when consistency matters.

anthropic/claude-haiku-4.5 · temperature 0.3 · max_tokens 1024.

Domain protocols

Optional. Named procedures the brain runs when input matches a domain. Lets one brain handle multiple input shapes without bloating the system prompt.

Domain: "research request" -> protocol: "Cite three sources, summarize each, then give a verdict." Output: "Markdown with H2 per source."

Visibility

Private brains are yours alone. Public brains land in the marketplace where other Brainboot users can install and run them. Make it public when the test cases pass and the system prompt has been edited at least three times.

Start private. Iterate until you trust the output. Flip to public.

Worked example

A complete brain, end-to-end

Five fields, one brain. Input on the left, output on the right.

The brain// editor.tldr

name

Editor TL;DR

system_prompt

You are a ruthless senior editor. Compress any input to a five-bullet TL;DR. Each bullet is one short sentence. Lead with the most surprising fact, not the topic. Never include throat-clearing.

execution_rules

Exactly five bullets, no more, no fewer.
No bullet exceeds 18 words.
No bullet starts with the word "The".

output_format

markdown

model

claude-haiku-4.5

temperature

0.4

max_tokens

512

What it produces// run_001

input

[pastes a 3,000 word essay about a power-grid failure in Texas]

output

* A single deregulation vote in 1999 made the rest of the failure inevitable.

* Texas runs an isolated grid specifically to escape federal weatherization rules.

* Wind turbines failed last; gas plants failed first by a five-to-one margin.

* Fatalities came mostly from carbon monoxide poisoning, not cold exposure.

* The grid operator's board members lived out of state during the event.

The same brain run on a different input produces five new bullets in the same shape. That predictability is what the rules buy you.

Quickstart

Three steps, blank to running

Open the editor

Hit the button below. The Brain Editor lives at /brains/editor. You'll land on a blank form with all five fields.

Fill three fields

Name, system prompt, and one execution rule. That's the floor. Everything else has a sensible default. Most authors stop here on their first save.

Run it from /chat

Save sends you back to your brain library. Pick the brain in the chat picker and send any input. If the output disappoints, edit the system prompt and rules - that loop is the work.

Open the editor

The first brain you write takes ten minutes. The hundredth takes ninety seconds. Start anywhere.

Free tier includes brain authoring. No credit card.