The short version
- A standard the model cannot see does not exist. Our playbook was written down and the batch missed it completely, because nobody put the file in the prompt.
- Show rather than describe. The same image request produced 2 of 3 usable frames from a written style description and 3 of 3 with one reference image attached.
- A brief moves the average of a batch without guaranteeing any post inside it. Anything a script can count belongs in the script that runs after generation.
AI output quality is decided before generation, in the brief. A batch of twenty posts we generated missed our own documented content playbook completely, because the playbook was sitting on a disk and not in the prompt.
Why does AI output miss a standard you already wrote down?
Because a generator returns the average of what it has read on the subject. A house playbook exists to be different from that average. Ours was a file nobody attached, which for this purpose is the same as not existing.
We found the problem by reading one of the posts against the playbook line by line and marking every rule it did not follow.
The same five inputs, briefed thin and briefed properly:
| Input | Thin brief | Loaded brief | What we saw |
|---|---|---|---|
| Style | Described in a sentence | One reference image attached | 2 of 3 on-style, then 3 of 3 |
| Standard | Named, not attached | The playbook file pasted into the prompt | A twenty-post batch missed the documented playbook entirely |
| Rules | Assumed to be known | Banned words and constructions written out | A rule restated in a doc does not survive generation |
| Facts | Model supplies the numbers | A fact file supplied, numbers only from it | Unsourced figures arrive fluent and wrong |
| Check | Read the draft | Count the rendered page | A draft will claim compliance the page does not have |
What does an AI brief carry that a human brief does not?
Hand a writer a thin brief and they come back with questions, or they open the standard because they know they will be judged against it. A model returns a draft either way.
So the brief carries everything a person would have gone and found. After that batch, ours stopped naming the playbook and started carrying it: the rule set pasted in rather than referenced, the banned constructions written out, one worked example that already clears the bar, and the fact file attached with an instruction that no number may come from anywhere else.
Why is showing better than describing?
We measured this where the result is easy to score. Describing the style in words produced 2 of 3 usable frames. Attaching one reference image to the same request on the same subject produced 3 of 3.
A written description competes with everything the model has read about the words inside it. A reference image is a target with nothing competing against it. The same holds in text: one post that already meets your bar beats a paragraph of adjectives about tone.
Where does the brief stop and the check start?
The brief moves the average and cannot guarantee any single output, so anything a script can verify belongs in a script.
We run both. The publish gate on this site counts the words that actually render rather than the words in a draft, and that check does not need a person.
The mechanical rules moved into the gate. What stayed in the brief is the part a script cannot check.
The model will not open your standard. If it is not in the brief, your standard does not exist.
Who should be writing the brief?
The most senior person on the account. This is awkward in a services business, because a client can see a draft and cannot see a brief.
Related: the priors no brief can argue with and the rules a script has to hold.
Frequently asked questions
Why does AI-generated content miss a company's style guide?
Because the guide is not in the prompt. A generator returns the average of what it has read about the topic, and a standard it cannot see carries no weight. We produced a batch of twenty posts that missed our own documented playbook for this reason.
What should an AI content brief include?
The standard as an attached file, one worked example that already meets the bar, the banned words and constructions, the source file every number must come from, and the check that will run afterwards. Anything a human writer would have gone and found has to be supplied.
Is it better to describe a style or show an example?
Show one. Describing the style in words produced 2 of 3 usable frames, and attaching a single reference image produced 3 of 3 on the same subject.
Can a good brief replace a review pass?
No. A brief raises the average quality of a batch without guaranteeing any single output. Rules that can be checked mechanically, such as a banned character or a missing table, belong in a validator that blocks publication.
Who should write the brief for AI-generated work?
The most senior person available. Specification is where the quality of a batch gets set, and a review pass can only catch what the brief already asked for.
How do you tell whether the brief or the model was the problem?
Read a failed output against your standard line by line and mark every miss. If each miss corresponds to something you never supplied, the brief was the problem. If a miss contradicts something you supplied, it is enforcement and belongs in a script.
Sources
- Chua Network delivery data across 8 client accounts (internal fact bank)
- Chua Network engagement records, anonymized (internal experience bank)
