Don't let the AI invent your metrics.

The most tempting shortcut in review season

You've got your rough notes. You open an AI chat, paste them in, and type some version of: "turn this into impressive performance-review bullets."

What comes back is genuinely good — tight, confident, well-structured. It also, quietly, contains a number you never gave it. "Reduced latency by roughly 40%." "Improved onboarding completion by about 15%." You didn't measure either of those. The model produced them because you asked it to make you sound impressive, and a plausible-sounding metric is what "impressive" looks like.

That's the moment worth talking about, because we spent a while building the opposite behavior on purpose — and the reasoning generalizes to any AI you point at high-stakes writing about yourself.

A made-up number isn't a flourish. It's a liability.

In most writing, an AI making something up a little is a nuisance you edit out. In a performance review or a promotion packet it's different — because someone is going to check.

Your manager was there. Your calibration peers were there. The number you can't source is the one a skip-level asks about in the room, and "the AI wrote that" is not an answer you want to give about your own accomplishments. A fabricated metric doesn't just risk being wrong; it undermines the true things next to it. Once one claim doesn't hold, the reader silently re-rates all of them.

So the fabrication that makes AI look helpful for career writing is exactly the failure mode that makes it dangerous for it. The impressive-sounding tool and the trustworthy tool are pulling in opposite directions.

The rule we actually wrote down

When we built the AI features in Cookd — a private tool for keeping career evidence — the first thing we wrote wasn't a prompt to make wins sound better. It was a constraint. Every AI flow in the product carries the same rule, in the system prompt, verbatim:

“Use only the information given. Never invent metrics, numbers, percentages, names, dates, or outcomes. If a figure is missing, ask for it — don't make one up.”

That's it. The interesting design work for high-stakes AI isn't deciding what the model is allowed to do. It's deciding what it's forbidden to do, and then routing the gap somewhere safe. When one of those flows can't find a number in your notes, it doesn't fill the blank — it leaves the metric empty and says what's missing: no measurable outcome here, add a number, a range, or a before-and-after. The feature's whole job, in that moment, is to admit it doesn't have the number and hand the work back to you.

That makes the AI less magical, and we decided the trade was the point. A tool that says "we need a number from you" demos worse than one that confidently writes "improved efficiency by 30%." But for a document you'll have to stand behind in a room, an assistant that flags the gap is worth more than one that papers over it.

The subtler version: the rubric is not your evidence

There's a second, quieter place the same failure shows up, and it's worth naming because it's easy to miss.

Feed an AI your company's leveling rubric — the description of what a "Senior" or "Staff" engineer is supposed to do — and ask it to write your case, and a helpful model will happily turn the rubric's language back into your accomplishments. "Demonstrates cross-org technical leadership," it writes, as if you'd told it you did. You didn't. That's the expectation, not the evidence.

We had to forbid that one explicitly too. The rubric text describes what the company wants at a level, so the model can use it to understand what each competency means — but it is never evidence of your own work, and it never gets quoted back as an accomplishment. The distinction between "here's the bar" and "here's what I did" is the entire game in a promotion case, and an AI that blurs them writes you a beautiful packet full of claims you can't back.

What this means for whatever AI you use

You don't need our tool to apply the lesson. The rule is portable.

Shaping is real help — tightening wording, fixing passive phrasing, turning a rambling note into a clean STAR bullet. Sourcing is where it turns on you: the model has no way to know which of your numbers are real, so it treats a made-up one and a true one exactly the same. If it produced a metric you can't point to, cut it. If it stated an outcome you can't defend, cut it.

And a fair limitation, since we're asking you to trust the honest version: no AI is perfect, ours included. Grounding the model in your own notes and forbidding fabrication reduces the failure; it doesn't eliminate it. The output is a strong first draft, not a finished truth — you're still the last check before anything goes in, which, for a document about your own career, is exactly where the check belongs.

That's also why the whole thing still works with the AI turned off. If the most valuable thing the assistant does is decline to invent your accomplishments, then the version with no model at all — the one that just keeps your real notes and asks you for the number — was never the degraded mode. It's the point.