Briefs

Supervision by description — articulate the standard and have labelled samples synthesized from it.

A brief states the standard a model should learn in structured natural language. It is supervision by description: you articulate what good looks like, and the build synthesizes labelled samples from it before training.

Definition

A brief is articulated intent — the standard expressed as words rather than examples. It suits the common starting point where you can describe the judgment you want but have no labelled dataset to hand. From a brief, the build generates the examples a model needs to train.

Mechanism

A brief that produces a usable model covers six elements:

Element	What it states
Intent	What the model is for, and the decision it supports.
Domain	The kind of content, its context, and its audience.
Exemplars of good	The qualities that should score well — described, not listed.
Failure modes	The anti-patterns and common ways content misses.
Edge cases	Tricky calls you want handled a particular way.
Constraints / non-goals	What the model should not optimize for.

At its most minimal, a brief is the essentials of a single trait — a pair of poles, or a question. The more of the six elements it carries, the closer the synthesized samples land to the standard you intend.

Interpretation

• A brief trades precision for speed. It is the fastest way to a first scoring model, and a strong starting point to refine with direct samples or feedback once real content flows.
• Describe qualities, not instances: “concrete and free of hedging” generalizes to new content; a list of ten good sentences does not.
• Name the failure modes explicitly. The negative pole shapes a trait’s geometry as much as the positive one, so a brief that only describes “good” leaves the boundary underspecified.

Edge cases

• A brief and supplied samples can coexist: synthesized examples fill the gaps a partial dataset leaves.
• Because a brief resolves to samples, a model built from one produces the same score card shape as a model built from data.

Next

Samples

What a brief is synthesized into.

Authoring schemas

Where a trait’s minimal specification is declared.