The Prompt Is the Spec

The Prompt Is the Spec

Ask most people what prompt engineering is and they describe wording tricks. Say please. Add “think step by step.” Try five phrasings and keep the one that felt best.

The two prompt changes AAA has shipped were not wording tricks. Each changed what the model had to produce. The fields that exist, what each one must contain, and which words are banned all moved into the prompt. That is a spec, and the prompt is where it lives.

Both changes live in one file, in the two prompts that generate every Active GEO and Active Pulse report. Both are quoted below unedited, straight from the commits that shipped them.

Diff One: A Voice Rule, Not a Style Wish

On July 29, 2026, a real Active Pulse summary came back with the word “However” in it. That word sits on the banned list every piece of AAA content is checked against, this one included.

The cause was simple. The prompt specified the JSON structure of the report and nothing else. It never said a word about tone. The model filled the gap with the corporate register it learned everywhere else.

The fix was a single constant, appended to the end of both prompts. Here is the live text, with the 31-word banned list trimmed for space.

Read the rules and you will recognize them. They are the same rules in AAA’s own writing guide. Condensed to one paragraph, they were handed to the model as a requirement. A style wish says “sound professional.” A spec says which 31 words are off limits.

Diff Two: Every Score Has to Say Why

On August 13, 2026, the feedback on real reports was “great, but could be tweaked.” Two problems sat behind that. Recommendations read like template lines that could apply to any company. Dimension scores were bare numbers with no explanation anywhere.

The fix again went into the spec, not the wording. Every dimension gained a why field. Every SEO finding gained a detail field. Every recommendation gained a related_to field that has to match a real finding label. The Active Eval piece covers that last constraint in full.

The example in the prompt changed from a bare number to a number with a reason. Both versions below are real, taken from the prompt before and after the commit.

Then a rule went under the example, and the rule is the part that does the work. Quoted exactly, punctuation included.

The example shows the model what good looks like. The rule tells it what good requires. Both are needed, and neither is a wording trick.

A Prompt Change Has a Bill

Richer output costs more tokens. The new fields add roughly 400 to 500 output tokens to a typical report. The output ceiling moved from 4,000 tokens to 6,000 in the same commit, as headroom rather than a hard need.

That line matters because prompt changes get treated as free. They are not. Every field added to a spec is paid for on every report, forever. Add the field anyway when it earns its place, and know the price.

How AAA Knew It Worked, and What It Did Not Check

Two checks ran. The related_to constraint runs on every report and fails any recommendation that does not point at a real finding. A person then ran real audits on a preview build and read every recommendation. Each one had to name something concrete about the actual site. They did.

What did not happen matters too. No held-out set of audits was scored before and after the change. The Active Eval piece describes that method. It was not applied here, and that is the first thing the next prompt change should fix.

Where Else This Pattern Applies

Writing the prompt as a spec generalizes past audit reports. None of the four examples below are real AAA client work. They show where the same move applies.

• Sales. A proposal drafter that must cite a real pricing-sheet line item for every cost, never a rounded figure.
• Finance. A variance commentary generator required to name the account and the dollar amount behind every sentence, in under 20 words.
• HR. A job description writer that must name the real team, the manager’s title, and one tool the role uses daily.
• Operations. A shift summary that must name the machine and the hour for every incident. Vague words like “issues” are banned.

Each example is illustrative, not an engagement. The two diffs above are the instances of this pattern that are real and running today.

What This Article Proves

This article was checked against the same 31-word banned list the prompt enforces on every report. Its PDF was generated and rendered to a real image before it shipped, using the Active Ship pipeline. That check caught a real defect in this article’s first PDF. A page break had dropped four bullets into the footer font. Fixed before it shipped.

Active Prompt is the name for this practice area. It means prompts written as specs, with the fields, the rules, and the bill in the open. Next up is Active Retrieval, which grounds answers in real documents and needs an eval to prove it.