@wulfie_bain_ - Why Top AI Startups

Open original

@wulfie_bain_ on X An OpenAI Startups applied AI lead argues that prompts in even the best AI companies grow contradictory over time because teams only add lines and never remove them. He proposes treating prompts like code and product decisions, using a modular, MECE structure (Background, Behaviour, Output).

  • Prompts evolve accretively, so contradictions and ambiguity pile up. Conditional prompts spread across files make the whole thing hard for anyone to review end to end.
  • He reports results from prompt rewrites: one startup cut agent speed-ups of about 50%, another saw 40% higher 7-day retention, and another cut costs by 30%.
  • Smarter models don't fix ambiguity. Someone still has to specify what the product should do, and evals force that clarity.

Prompts at top AI startups tend to rot in a predictable way. Teams only ever add to them. Someone patches a bad behavior by appending a few lines, nobody reads the whole thing end to end, and six months later the prompt contradicts itself in three places. Add conditional prompt injection, where different engineers own different files that get stitched together per scenario, and you end up with thousands of lines of spaghetti where the first paragraph is forgotten by the time the reader reaches the last. The author, Wulfie Bain, who leads applied AI engineering for OpenAI's startups team across EMEA and APAC, says this shows up even at companies with hundreds of millions in ARR. He also says the problem is ambiguity as much as contradiction: an instruction like "never refer to competitors" makes perfect sense to an engineer who thinks about competitors all day, but tells an agent nothing about who counts as a competitor or what to do with a partner you sometimes collaborate with. His fix rests on two principles. First, treat prompt changes as product changes, because agent behavior is the product. Formatting, response speed, whether the agent asks permission before acting, whether it replies in the user's language: these are UX decisions, and whoever writes the prompt is making them whether they realize it or not. If the team can't say what the agent should do in a given error trace, the model can't be expected to do it either. Second, treat prompts as code: modularize them, keep the sections mutually exclusive and collectively exhaustive (the MECE idea he borrows from consulting), refactor them when they bloat, and maintain them actively. He reports concrete results from rewriting prompts this way over a few days of work: one startup sped up its agents by 50%, another raised seven-day retention by 40%, another cut costs by 30%, with gains holding across models from different providers. He cites a Lovable engineer who reportedly cut LLM spend by $20M a year after his mother spotted duplications in a prompt. The practical payoff he claims is faster iteration, fewer regressions, and better agents, and he proposes a simple, MECE-style prompt template at the end of the piece. The core of his argument is that clear, specific, non-redundant instructions are a product discipline, not a writing style. A reader who only takes one thing from this should take the habit of reading the whole prompt cold, the way someone with no context would, and asking what each line actually requires.

Ask about this article