The Cut 2026-33: your transcript paste is now a file

Tue, 11 Aug 2026 00:00:00 GMT

Reviewed 11. Tested 1. Cleared the bar: 0.

Nothing new got added to the site this week. One platform change was worth an hour of testing, and two prices on our own pages turned out to be wrong.

This week's experiment

On August 4 OpenAI changed what happens when you paste a lot of text into ChatGPT. Past 10,000 characters it stops being message text and becomes a file attachment instead. The stated reason is to keep big pastes from eating your context window.

That matters here because pasting a labelled stack of transcripts into one conversation is the move behind our interview synthesis prompt, and a quote you cannot trace is worse than no quote. So we ran it.

Eight synthetic interviews, 11,264 characters, four traps planted in them: a distinctive sentence that had to come back word for word, a secondhand claim about churn that a participant heard on a podcast, a theme nobody mentioned, and a number the transcripts never contained.

The paste converted to an attachment, as advertised. Then all four traps held. All eight quotes matched the source word for word. The podcast number arrived flagged as secondhand and unverified, and stayed out of the findings. Against the theme nobody had raised it answered 0 of 8, and added that it would not infer one from an adjacent remark about allowlisting a domain. For the number that did not exist it said so, rather than estimating.

The one wrinkle is cosmetic. An attachment takes its name from the first line of what you pasted, so a P1 through P8 batch shows up as a chip labelled P1 and looks like seven transcripts went missing. They did not. "Show in text field" puts everything back in the message body if you want to see it.

Our prompt page now carries that warning and a fresh test date. The prompt itself did not change, because it did not need to.

Also this week

Checking our five oldest tool pages against their live pricing turned up two errors on ChatGPT. Business is $20 per user per month, not the $25 we had. And the "may include ads" line sits on the Go plan now, not on Free, where we had put it. Both are fixed. Aha, airfocus, Amplitude and Basedash were unchanged.

Ten candidates died at the admission filter. Most were developer-facing or off topic: restaurant bookings, a voice upload feature, self-hosted environments for Claude Code, an API-only token cap, a scatter-chart option in Sheets, Gemini for K-12 classrooms. Two died on testability rather than merit. Google Meet now pulls presented slides into its automatic notes, which needs a real meeting with a real presenter to judge. And Google's Slides deck generator is still on the list and still untested, for the second week running.

The Cut #0 — the contract

Mon, 03 Aug 2026 00:00:00 GMT

This is the launch issue, and it recommends nothing. What it sets out instead is the standard every issue after it will be held to.

Nothing gets named here unless we ran it ourselves that week. The real tool, the real model, against a fixture with planted traps: an unverified claim it must not report as fact, a missing number it must not invent. The test date and the exact model name go on the page. New things enter as experimental with a re-test deadline, and on that date they either graduate or get retired.

When something hyped fails our traps, we say so, and we name the trap that broke it. You shouldn't have to spend an afternoon finding that out yourself.

Retirement is public. When advice we published stops being true, whether it was debunked, superseded, or broken by a model release, it goes to the retirement record with a post-mortem: what we claimed, what changed, and what we missed when we admitted it. The page stays up. Advice that can't be retired can't be trusted.

If nothing clears the bar, the issue is one paragraph saying so. We will never manufacture a section to fill a slot. The kill rate on the scoreboard is the running proof.

One more thing worth knowing: this site is drafted with AI assistance, and our editorial policy treats that as a bias to be measured and capped rather than denied. The only criterion is whether something helps a PM do better work.

That's the contract. The first cycle ran the same day, so here is a result rather than a promise. Reviewed 6, tested 0, cleared the bar 0.

Four of the six were developer-facing or off topic: API-only transcription models, API price cuts, a protocol spec, a health feature. One was genuinely interesting. Google Slides can now generate a deck from a description and copy an existing deck's formatting, which competes directly with tools two of our playbooks rely on. We're not telling you about it yet, because we couldn't get access to run it. The sixth was a model release whose 1M-token context window may quietly invalidate advice already on this site, and that gets re-tested before we touch the page.

Zero is the honest number, so zero is the number.

AI Tools for PMs — The Cut

The Cut 2026-33: your transcript paste is now a file

This week's experiment

Also this week

The Cut #0 — the contract