Two Viral Posts Prove the AI Bottleneck Isn't the Technology — It's You.
Two Hacker News posts, a combined 1,374 points, same week. One named the upstream failure — vague requirements in, vague output out. The other named the downstream one — raw AI text pasted straight into Slack and email. The model was fine in both cases.
By Forge Team
The AI output quality problem most professionals complain about is not a model problem. Two posts that hit Hacker News in the same week — a combined 1,374 points — showed where the actual failures sit. One is upstream: the brief going in is too vague, so the output is too vague. The other is downstream: the output comes back and gets sent to colleagues without editing. Fix one end and not the other, and you've halved the problem at best.
What happened
On May 17, software engineer Frederick Vanbrabant published "AI Won't Speed Up Your Processes" — 678 points on Hacker News. His argument: AI cannot fix a slow process because the bottleneck is never the model. It's always upstream — vague requirements, incomplete documentation, problems that haven't been defined clearly enough to hand off to a person, let alone a language model. He estimated that the fast-build-with-AI stories routinely omit the 40-plus days of specification work that actually gate the project. The model isn't slow. The brief is.
Four days later, "No Slop Grenade" reached 696 points. It named the downstream failure: the growing pattern of pasting AI-generated text directly into Slack messages and client emails without editing them. The complaint was blunt — colleagues are receiving walls of generated text that they now have to compress and verify themselves, doing the work the sender skipped.
Simon Willison flagged a third version the same week (simonwillison.net, May 24): Armin Ronacher's critique of AI-generated reports that sound authoritative but are speculative — confident tone, guessed diagnosis. The output reads like it knows. Often, it doesn't.
Three distinct failure modes, same underlying pattern: treating the model as both the input processor and the output quality gate. It's neither.
What to change on Monday
Two fixes. One for each end.
Upstream: before you prompt, write a brief. Not a long one — three lines is enough. Who the output is for, what decision or action it needs to support, and what format it should take. Vanbrabant's point isn't that AI is useless for complex problems; it's that a vague prompt amplifies whatever ambiguity is already in the process. A tighter brief reduces that at the source.
Downstream: before you send, edit. The three-second test: would you send this exact text if you'd written it yourself? If not — if you'd cut the third paragraph, tighten the opening, or remove the sentence that says "it is worth noting that" — the text isn't ready. Editing AI output is a writing skill, not a sign the model failed.
Dara: the upstream problem
Dara is the content and campaigns lead at a 45-person SaaS. She uses Claude to draft competitive comparison pages. Before she started scoping explicitly, the output was generically accurate but not useful — correct facts, no specific differentiation claims, no language tuned to the technical buyer evaluating procurement risk.
Her fix was simple. Before every competitive brief she writes three lines: the buyer role, the specific objection the page needs to handle, and the competitor claim she's countering. The output stopped being generic the moment the input stopped being generic. She didn't change models. She changed what she handed to the model.
Practice writing a brief that gives the model enough to work from before you run the prompt.
Kwame: the downstream problem
Kwame is an operations lead at a 90-person consulting firm. He sends weekly status updates to three clients. For six weeks he pasted AI-generated summaries directly into client emails without editing them. Then a client flagged one: the summary described a project risk as "being actively monitored" — a phrase the model produced, not a fact Kwame had verified or even intended.
The update sounded authoritative. It wasn't. He now does two passes on every AI draft before it leaves his outbox: one to verify any factual claims, one to cut anything he couldn't personally defend if asked. It adds five minutes per update. It removes the version of him that sends things he hasn't actually read.
Take a raw AI draft and edit it down to something you'd actually send.
The model isn't the problem
A better model given a vague brief produces a better-sounding version of the wrong thing. The upstream work — defining what you need clearly enough to prompt for it — and the downstream work — editing output before it reaches another person — are still yours.
Learn to recognise the patterns in AI output that signal the model was generating rather than knowing.
Like this post?
Get the next one in your inbox. Practical AI skills, no filler.