Teams blame model quality for output that was doomed by the instruction. A brief that says write a blog post about WhatsApp marketing produces exactly what you would expect: a summary of what is already on page one.
What a usable brief contains
- The specific query the piece targets, not the topic
- Who the reader is and what they already know
- The material only you have: your data, transcripts, screenshots, internal numbers
- What the piece must not say, including claims you cannot verify
- Two examples of the tone you want, pasted in full
- The structural shape: how it opens, what order the argument runs
The material clause does the heavy lifting
A model given your customer call transcripts writes something that could not exist elsewhere. A model given nothing writes an average of everything already published. Same model, entirely different outcome.
The must-not-say clause
On regulatory topics this is not optional. A model will state statutory positions confidently and be out of date. Naming what it must not assert, and supplying the current text yourself, is what makes the output publishable.
Judge the brief, not the draft
When output disappoints, reread the instruction before changing the model. If the brief did not contain anything the model could not have guessed, the draft was never going to be worth publishing.
General information, not legal advice. Rules current as of July 2026.