As promised: The six-section brief that replaced every long chat prompt I wrote

Long prompts decay.

You keep stuffing context into one chat window until the model is drowning in your own mess.

The fix is not a better prompt. It is a brief written to disk.

A brief is a contract. Any worker opens it cold, executes, hands back a result. No shared history. No "as we discussed earlier." The doc is the entire relationship.

Here is the template I use. Six sections. Every brief has them. In this order.

# <slug>

cd <absolute path the worker starts in>

## Task

One paragraph. The outcome. What "done" looks like in prose.

## Context

Paths to files the worker needs. The spec. The design doc.

The prior handoff. Anything not obvious from the codebase.

## Scope

- In-scope: what the worker is allowed to touch

- Out-of-scope: what the worker must not touch

- Kill-switches: time or complexity thresholds that trigger a fallback

## Acceptance

- Binary bullets. Either true or false.

- Test counts. Commit messages. Evidence file path.

- No subjective language. "Looks good" is not acceptance.

## Escalation

One file path. If the worker hits ambiguity, it writes a question

to this file and exits. Main session reads it on the next turn.

## Returns-with

- Commit SHAs

- Test pass count before and after

- Wall-clock per major task

- Final handoff doc path

Why each section load-bears

Task forces you to state the outcome in prose before a single line of code exists. If you cannot state it, the work is not ready to dispatch.

Context is the line between a clean run and a hallucinated one. Never rely on the worker guessing which file you meant. Name the paths.

Scope stops the worker sliding into adjacent work. In-scope is the yes list. Out-of-scope is the no list. Both matter. Kill-switches protect you from a worker burning six hours on the wrong branch.

Acceptance is the section most people skip. Binary only. If you cannot write acceptance criteria, the work is not shaped yet. You are not ready to brief.

Escalation is the pressure release valve. Without it, a stuck worker invents nonsense. With it, the worker writes a question file and exits clean. You read it on your next turn. No wasted tokens.

Returns-with is the receipt. It is also the input for the next brief. Every worker writes a handoff doc on the way out. That doc is what the next worker reads in.

The cold start rule

Every brief is written assuming the reader has seen nothing before.

No shared session.

No "continuing from yesterday."

No implicit knowledge.

If the brief cannot stand alone, it is not finished.

This one rule is why the whole system works. A brief that survives cold start also survives parallel workers, swapped models, and weeks-later continuation.

The principle

You do not prompt a worker. You contract one.

The brief is the interface. The worker is interchangeable. The doc is the system.

Give this MD file attached to you chosen LLM and it will guide you though set up.

// A<3

13 comments