My Portfolio Read Like a Compliance Audit
In July I had agents write the showcase cards for my own projects: fourteen cards, about 150 prose fields, each one built from a verified evidence ledger so nothing on the page could claim something the repos didn't support.
The result was accurate and I couldn't stand to read it. It took an eighteen-agent diagnostic pass to tell me why, and the answer is a failure mode in model-written prose I haven't seen described anywhere.
The Grammatical Subject Was a Prior Claim
Across 150 prose fields there was exactly one "I," and it was inside a quotation of a system speaking. Nothing anywhere said that I built, chose, or decided anything.
The cards were written as a re-audit report addressed to a reader holding the previous version of the document. The grammatical subject of nearly every sentence was a prior claim, a commit, or the harvest run that produced the evidence, never me and never the system. The frame was a person and the contents were an asset register.
That's what happens when the writer's job is defined as "verify the ledger." The writer verifies the ledger and its sentences are about the ledger, so a reader who wasn't holding the ledger has no idea what any of it is for.
The Negation Rhythm
The second finding was a cadence. The same move landed on eight of fourteen cards in a row: not X, as previously claimed, but Y. Test-pinned rather than demo-only. Not a four-service monorepo. Not the figure in the earlier draft.
Read one and it sounds careful. Read eight in a row and the page starts to sound like a document defending itself, with scare quotes around the thing being corrected and the audit holding a claim at arm's length. The reader starts to wonder what the argument was, and there wasn't one. There was a verification pass, writing down that it had verified.
The diagnosis split those into two kinds that are built exactly the same way. "Test-pinned rather than demo-only" is the good voice: it tells you a real property of the thing. "Not a four-service monorepo as previously claimed" is the audit voice: it tells you about a document you haven't seen. Keep the first and cut the second. They look the same on the page and they're opposites.
The Number That Disagreed With Itself
One card's summary said a project had the second-highest commit count in the portfolio, and gave the number. Its own chip, four lines below, gave a different number, seven higher. Both were tagged as measured. The metric pipeline takes the newest figure, and the prose had baked in the one it replaced.
It's a small defect, and it's the whole problem in miniature. The prose was true when written and the chip was true now. Nothing connected them, so the page carried both, and a reader who checks one thing checks that.
What the Agents Found and I Did Not
I couldn't name any of this before the diagnostic pass. I knew the page felt wrong, and I described it as "I can't pinpoint it." Five independent lenses read the cards, three competing proposals for a new card anatomy were built, nine judges scored them, and one write-up came back with the sentences above.
What I couldn't see was that the cards had no author. Every sentence was defensible and no sentence was mine. A portfolio is the one document where that's disqualifying. The reader is trying to find out who you are, and the page was answering a different question: whether the previous page had been accurate.
The Fix
The rewrite has a subject. Sometimes it's me, sometimes it's the system, and it's never the ledger, the commit, or the prior draft. The verification still happens, in the same evidence file, with the same tags. It just doesn't get to write the copy.
The numbers in the prose are placeholders now, filled in at build time from the same source as the chips, so a sentence can't disagree with the number next to it. That build deserves its own writeup. This is where I first needed it.
If you have a model writing about your work from an evidence base, count the first-person pronouns. If the answer is zero, the model wrote an audit, and you're the thing being audited.
-- Justin Higgins. Software Engineer, Midwest. Had agents write his portfolio, found one "I" in 150 fields, and gave the sentences a subject.
Companion pieces: The Work Looks Like Nothing - what the page was supposed to show.
Reactions, disagreements, war stories: jchigg2000.dev@gmail.com