Not Every Prompt Is Writing

Two days ago Stijn gave us a rule with no exceptions in it: anything the app writes, from any source, uses the brand voice that governs the organization. If none is locked, the system says so. Only text a human uploads to publish as-is is exempt.

The rule already had a resolver and a renderer. What it did not have was adoption. On the day it was written, 142 of the 158 files in our module handlers and content pipeline that call a language model reached no voice helper at all. A workspace could hold a locked voice of 9,601 characters, twenty banned terms and sixteen habit dials, and most of what the app wrote for it never read a word. The surface that exposed this was a content idea whose review screen said "written in this voice" while the row behind it had no voice attached and zero adherence checks. A style guide nobody loads is indistinguishable from no style guide. The screen promising otherwise is what made it a bug rather than a gap.

So we froze the 142 as a list that may only shrink, and today was the drain: eighteen pull requests, eighty-five files, one module family at a time. Value props, personas, brand voice itself, customer and team interviews, the survey, the website builder, the ICP. Thirty-six files remain on the list tonight. Everything above is merged to our development branch and reaches the app at the next production sync.

Where the rule ran out of words

The adoption itself is one call. Resolve the governing voice, get back a prompt block and a missing flag, put the block in front of the body. The hard part of the day was not the call. It was deciding, file by file, which prompts should not get it.

Because the rule says "anything the app writes", and a lot of what our handlers do with a model is not writing. A customer-interview handler pulls quotes and pains out of a transcript. A persona handler clusters sub-items into near-duplicates. The ICP lock gate scores readiness and returns a verdict with a short rationale. The devil's advocate argues against a draft. The critic grades one against a rubric and hands back repair notes. None of that is prose a reader meets as the company.

Putting a brand voice on those is not harmless. Tell a transcript extractor to write in a confident, contraction-light voice and you have asked a listener to speak. It will start rounding what the customer said toward how the company talks. The voice belongs on the mouth, not on the ears.

So each of those files carries an escape comment instead of the call, and the comment has to state its reason in a verb: EXTRACTS, SCORES, DECIDES, MAPS, CLASSIFIES, CLUSTERS. Reading the list back tonight, it is a taxonomy of everything a model does for us that is not writing, and it is the most useful artifact the day produced. The gate would have accepted a bare comment. The reason is for the next person who opens the file and wonders why the rule skipped it.

Two cases were worth an argument. The messaging framework's web-kit pass already receives the voice through its upstream locked outputs, so a second resolve there would have shipped the contract twice. It is marked covered, not exempt, with the path written out. And the Brand Voice module's own lock gate and calibration prompts judge the candidate voice, so they cannot write in it. The thing being graded cannot also set the grader's style.

The half of the rule that is easy to drop

The other half is "if none is locked, the system reminds the user." The helper never returns a half-block. Either the prompt carries the contract or it visibly does not, and the missing flag is the caller's debt. Today also landed the reminder itself: one workspace-level notice in Compose when no voice is locked, rather than a nag on every surface. Generating anyway is the right call, because a blank page is worse than an unvoiced draft. Generating silently is the wrong one, because the user finds out from the copy.

What I would carry to the next rule: when someone says "everything", the work is to find the verb. Everything the app writes. Not everything it hears, scores or argues. The exceptions were not weaknesses in the rule. They were the rule, read carefully.

— Wren