Anthropic's own developers say better AI results start with your blind spots, not longer prompts
An Anthropic developer shares techniques for working with Fable 5 — and reveals the company cut 80 percent of Claude Code's system prompt because newer models want less instruction.
If your AI results keep coming out almost-but-not-quite right, the problem may no longer be the AI. Anthropic developer Thariq Shihipar argues that with Claude’s newest model, Fable 5, output quality is mostly limited by something the model can’t see: the things you didn’t think to mention.
Shihipar sorts what you bring to a prompt into four buckets. Known knowns are what’s in your prompt. Known unknowns are the open questions you’re aware of. Unknown knowns are things so obvious to you that you’d never write them down. And the dangerous one: unknown unknowns — the considerations you haven’t thought of at all. His observation: being too specific is as bad as being too vague. Overly detailed prompts lock the model into a flawed plan; vague ones get you generic industry defaults that don’t fit your case.
His fix is to use the AI to find your own gaps before the real work starts. In a “blindspot pass,” you literally ask Claude to identify what you’re missing — for example: “I know nothing about the auth modules in this codebase. Can you do a blindspot pass to help me prompt you better?” He also recommends structured interviews (Claude asks you questions one by one, starting with the ones that would change the whole approach) and quick brainstorming rounds with several radically different prototypes to react to.
What’s behind it: this isn’t just one developer’s workflow — it reflects a shift in how Anthropic itself steers its models. In a recent conference talk, Shihipar revealed the company cut 80 percent of Claude Code’s system prompt (the standing instructions the tool feeds the model). Newer models “want a smaller system prompt,” he said, and examples “tend to constrain it because it’s actually more imaginative than the examples we give it.” For years, the advice was more instructions, more examples. That era seems to be ending: the models fill gaps well — as long as you’ve told them where you actually stand.
What this means for you: If you’re new to AI tools, this is encouraging — you don’t need to master elaborate prompt formulas. Say where you’re starting from (“I’ve never done this before, ask me what you need to know”) and let the model interview you. If you’re a power user, try the blindspot pass on your next tricky task and consider trimming your own carefully hoarded mega-prompts. A fair caveat: this guidance is tuned to Anthropic’s newest models; older or smaller models still benefit from more explicit instruction.
Sources
Source: https://thariqs.github.io/html-effectiveness/unknowns/
Researchers made AI models run a startup for 500 days. Most went bankrupt.
Princeton's CEO-Bench tests whether AI agents can steer a company over time, not just finish tasks. Only three models ended above their starting capital — and a dumb rule-based script beat almost everyone.
