[{"content":"I bet you haven’t read your system prompt.\nNot skimmed it.\nNot had it summarised for you.\nActually read it.\nThe whole thing.\nEvery token.\nAs the LLM sees it.\nWith jinja templating rendered, with AGENTS.md appended, with few-shot examples, with tool schemas. Everything.\nTry it. You’ll doubtless find dozens of little issues. Poorly explained concepts. Overly verbose language. Redundant information. Inconsistencies between prompt and tool specs.\nIt’s wild when you consider how important the system prompt is, and how little effort it takes to read it, and yet no one really knows what’s in it.\nIt really isn’t even that hard. Back of the envelope: a pretty big system prompt, with few-shot examples, and a bunch of tools might be 10k tokens. 0.75 words per token means it’s about 7.5k words. Reading speed is about 250 words per minute. So half an hour. Half an hour for you to read every single word. That’s nothing in the grand scheme of things.\nEvery time I have done this, I have come away with a much better, cleaner prompt, lower costs and a quality bump.\nTry it out. Get your agent to render your system prompt as it would appear at runtime. And then sit down and read it. All of it.","date":"31 Aug 2026","permalink":"/posts/unreasonable-effectiveness-of-reading-your-system-prompt/","section":"posts","summary":"I bet you haven't read it.","title":"The Unreasonable Effectiveness of Reading Your System Prompt"}]