Tell agents the why, not just the how

Open original

Modern AI models are capable enough to make meaningful decisions about how to solve problems, so you should tell them your priorities and context instead of just giving them a narrow spec. This lets them suggest better approaches and avoid wrong assumptions about what you actually want.

  • Early AI agents needed explicit step-by-step instructions; now they fail because they misunderstand your goals, not because they're confused about how to execute
  • Sharing broad context—your long-term aims, constraints, and what tradeoffs matter—lets models suggest improvements you wouldn't have thought to specify
  • Explicitly ranking your priorities (e.g., "I care less about performance than observability here") gives models the information they need to make smarter choices

Early AI agents were basically useless without explicit step-by-step instructions—you'd have to tell them exactly which methods to add to which classes or they'd veer off completely wrong. That's changed now. Modern frontier models like GPT-6-Astra are competent enough that when they screw up, it's usually because they've made wrong assumptions about what you actually want, not because they're confused. The author noticed this with code generation: the model will happily produce minified code if it thinks it's writing for itself, but it's perfectly capable of writing readable code once you clarify that humans will be reading it. The real insight here is that you should tell agents *why* you're asking them to do something, not just *what* to do. The author's example prompt to his AI agent Deckard spends about half its length explaining the broader context: he wants to build a tool to detect and hide AI content, he's constrained by his MacBook's battery and heat, and he's personally invested in the project. This context lets the model make better calls—it suggested using native messaging and a different model than what the author had initially guessed at. For work, he does the same thing but emphasizes his technical priorities: which matters more, avoiding bugs or performance? Can observability be traded off for elegance? Models are smart enough now to actually use this information meaningfully. The mistake most people make is treating AI agents like they're search engines—firing off a concrete technical spec and expecting good results. That's the XY problem in action: you're asking for expert advice without giving the expert enough context to actually help. The author's pushback here is subtle but important: there's no magic prompt formula that works everywhere, and chasing perfect wording is missing the point. What matters is treating the model like a collaborator who needs to understand your actual goals and constraints, not just your immediate task.

Ask about this article