Prizerv

The AI quoting itself back as the user's own words

Some reports displayed a quote in the user's own voice that the user never actually said, at least one of them was the product's own prior sentence, laundered through the pipeline and handed back as someone's private insight. The instinct was to fix it in the prompt, tell the model to only quote verbatim, but a rule like that had already been silently ignored more than once on the same codebase before it was taken seriously. The real fix was a byte-exact checker: code that only trusts a quote if it can find those exact words, unparaphrased, inside what the person actually typed. Prompt instructions and a hard structural check are different classes of defence, and for anything attributed to a real person, only the second one holds.

The test method that quietly broke the thing it was testing

For a while, testing meant pasting a whole fake conversation into one message box and letting the app process it, the fastest way to generate test data. The problem: the app stored that entire paste as if it were one message from the user, meaning the fake half written by the AI itself sat inside a row labelled as the user's own words. That meant the same verbatim-quote checker above certified the product's own writing as real user speech, because structurally, that is exactly what it looked like. Caught by building a proper test harness that writes fake conversations directly and correctly, then re-running the checker against both versions of the same content and watching it flip the right way. Your test data generation method can be the exact source of the bug you are trying to catch.

Why the sessions sounded like a survey

The first version fed the psychologist's own design document straight into the model as its live instructions, 7,000-plus words of structured tables covering per-turn choreography, scoring rubrics, and question menus. It ran that way for about six weeks. The sessions worked, technically, but read like a form: evaluative openers, menus of feelings, one construct after another in the document's own marching order. The report generator, a single free-form generation with no such structure, already sounded genuinely good, which proved the model itself was not the problem. The fix was a full rewrite of the runtime prompt: no zones, no menus, follow the person instead of the agenda, and it was not a straight cut, an early draft was tested offline at under half the eventual length, then grew back to nearly double that once real testing surfaced what the minimal version was missing. The same swap also broke report generation for a day, since both had quietly been sharing one prompt function, caught by comparing report length before and after, and fixed by splitting the two apart for good. Sounding human was the hardest thing in the build, harder than any of the plumbing. A self-discovery tool lives or dies on whether the output feels genuinely personal rather than generic AI output, which is the pitch a lot of AI products end up having to make. More than twenty prompt iterations went into it and none of them were really about the model.

The build with the least obvious buyer

This is the most ambitious thing I have built and the one I would find hardest to sell. The three-phase session is genuinely useful, and the market for it is thin, badly defined, and crowded at the cheap end by personality quizzes promising the same thing for free. I could argue myself into a route to market, but I would be arguing. So I will say it plainly: this is the one I would struggle to sell.