Wulfie Bain, who leads Applied AI Engineering for OpenAI’s startups team across EMEA and APAC, spent a long-haul flight writing up a pattern he says he sees inside nearly every fast-growing AI company he advises: prompts that no longer make logical sense.
His claim, posted to X on September 10, is specific. It is not that these companies write ugly prose. It is that their instructions to their own agents contain contradictions and unresolved ambiguity, and that those defects show up directly in the mistakes the agents make.
Bain’s explanation for how this happens tracks a familiar shape from software engineering. A prompt starts as a short, clean document written by one person. As the product grows, engineers append fixes and edge cases rather than rewriting the whole thing. Nobody re-reads it end to end. Old instructions that contradict the new ones never get removed. The file grows until, by his description, a human reviewer forgets the opening lines before reaching the close.
A second failure mode compounds it: implicit knowledge. An instruction like telling a model to avoid mentioning competitors reads as obvious to the engineer who lives inside that company’s competitive landscape every day. To a model with none of that context, “competitors” is undefined. Bain says this gap, between what a team means and what they actually wrote down, is where most of the ambiguity hides.
Bain’s fix borrows two ideas wholesale from software practice rather than prompt-craft. First, treat every prompting choice as a product decision: whether an agent asks for approval before acting, whether it replies in bullet points, whether it searches the web by default. Second, treat the prompt file itself as a codebase, organized so that no two sections overlap and every section is reviewable on its own without re-reading the whole document.
Rather than one long list, his proposed template splits an agent’s instructions into three separate concerns that never bleed into each other: the world the agent needs to know before it runs, the internal logic it follows while it works, and the format of what a user actually sees at the end. Keeping those three isolated, he argues, means a single change (say, switching output from markdown to XML) touches one section instead of triggering side effects across the whole file.
The numbers Bain attaches (agents sped up by as much as 50%, retention gains as high as 40%, cost cuts as high as 30% after a rewrite) are his own account of client engagements, not independently audited results, and he presents them that way rather than as a published study. He also points to a Lovable engineer’s account of cutting $20 million a year in LLM spend after a family member spotted duplicated instructions in the prompt, another self-reported anecdote rather than a verified figure.
The underlying argument survives even if the specific percentages don’t hold up under audit: as agent products scale past a single founding engineer, nobody owns the prompt the way someone owns a codebase, and that ownership gap is where cost and reliability leak out. Any team running a production agent with more than one contributor to its system prompt should treat a full read-through, sections split by concern, as a recurring maintenance task rather than a one-time fix.
Wulfie Bain wrote this analysis in a post on X published September 10, 2026.
