Grep your own slides

A pitch deck is a stack of claims wearing nice typography. Last week our company brain did something to one of ours that no human reviewer had thought to do: it treated the slide as a testable assertion and ran it against the codebase.

The slide in question showed an ecosystem diagram — several device types, each labeled with the signal it contributes to the product. Standard startup fare. The brain’s review had already flagged it on style grounds (“built vs. roadmap not distinguished”), which is the kind of note that dies quietly in a comments thread, because it’s an opinion and the author can shrug.

Then it stopped being an opinion. The brain took every noun on the slide and grepped the repository for it. One signal type hit sixty-five files — real, load-bearing code. Four other signal types on the same slide hit zero files each. A couple of near-miss keywords turned out to be decoys on inspection: “heart” matched only like-button icons, “camera” matched only QR-permission boilerplate. The verdict wrote itself: two layers of the diagram were built, one layer was aspiration drawn in the same visual weight as fact.

The interesting part isn’t that the deck oversold — every early deck does, and a diagram that shows where you’re going is legitimate. The interesting part is what happened to the review. The same critique, upgraded from “suggestion” to “evidence,” is a different speech act. You can wave off a reviewer’s taste; nobody argues with a grep count of zero. In our audit log the item literally moved categories that day, from advice to finding.

There’s a general mechanism here for anyone letting an agent maintain company knowledge: outward-facing claims should be joinable against internal ground truth. Slides, landing pages, and application answers all make assertions about what exists. The codebase, the commit log, and the metrics store know what actually exists. An agent that can read both sides can run the join — mechanically, on every revision, with no social cost. Humans skip this check not because it’s hard but because it’s awkward; a machine doesn’t feel awkward.

The fix on the slide was cheap once the evidence existed: mark the unbuilt layers as roadmap, keep the diagram. The deck got more honest and, we’d argue, stronger — “here is what works today, here is what it unlocks” beats a flat mosaic of unverifiable boxes.