{"version":"v1","site":{"name":"expectedwrong","url":"https://expectedwrong.com"},"links":{"collection":"https://expectedwrong.com/api/public/posts","rss":"https://expectedwrong.com/rss.xml","llms":"https://expectedwrong.com/llms.txt"},"post":{"slug":"cognition-in-the-prompt","title":"The Cognition Is in the Prompt","subtitle":"Parahelp's six-page system prompt is less a set of instructions and more a blueprint for a mind.","url":"https://expectedwrong.com/cognition-in-the-prompt","api_url":"https://expectedwrong.com/api/public/posts/cognition-in-the-prompt","published_at":1748692800,"published_at_iso":"2025-05-31T12:00:00.000Z","updated_at":1771555422,"updated_at_iso":"2026-02-20T02:43:42.000Z","tags":["agents","prompting","llm","customer-support","design"],"excerpt":"Parahelp's six-page system prompt is less a set of instructions and more a blueprint for a mind.","meta_description":"Parahelp's six-page system prompt is less a set of instructions and more a blueprint for a mind.","reading_time_minutes":2,"word_count":329,"engagement":{"signals":0,"counterpoints":0},"body_markdown":"Parahelp ships an AI customer support agent, and their system prompt is six pages long.\n\nNot six pages of rules. Not a list of dos and don'ts bolted onto a base model. Six pages of *structural cognition* — the shape of how the agent is supposed to think, reason, escalate, and hold context across a support conversation.\n\nThis is the thing people keep getting wrong about agent design. The assumption is that the \"real\" architecture lives in the code — the tool calls, the memory retrieval, the orchestration layer — and the system prompt is the part where you say \"be friendly and don't make up prices.\" That framing is backwards.\n\nThe system prompt *is* the architecture. It's where you define what the agent notices, what it treats as a trigger for escalation, how it models the person it's talking to, what it considers resolved versus open, what it does when two priorities conflict. All the things you'd normally call \"cognitive structure\" — they live in the prompt, expressed in natural language, because that's the interface.\n\nA six-page prompt isn't bloat. It's the engineering surface.\n\nWhat makes Parahelp's approach worth studying isn't the length — it's that they apparently understood early that you can't outsource the hard thinking to the model's priors. A general-purpose model has general-purpose instincts, and customer support is a domain with very specific failure modes: over-promising, under-escalating, resolving tickets that aren't resolved, treating a frustrated user like a ticket number. You have to write that knowledge in. Explicitly. In the prompt.\n\nThe video and the blog post are worth the time if you're building anything agent-shaped. Not because the techniques are exotic — they're not — but because seeing someone actually commit to the prompt as a design artifact, rather than a config file you fill out at the end, reorients how you think about the whole thing.\n\nThe cognition doesn't emerge from the context window on its own. You have to put it there.","body_text":"Parahelp ships an AI customer support agent, and their system prompt is six pages long. Not six pages of rules. Not a list of dos and don'ts bolted onto a base model. Six pages of structural cognition — the shape of how the agent is supposed to think, reason, escalate, and hold context across a support conversation. This is the thing people keep getting wrong about agent design. The assumption is that the \"real\" architecture lives in the code — the tool calls, the memory retrieval, the orchestration layer — and the system prompt is the part where you say \"be friendly and don't make up prices.\" That framing is backwards. The system prompt is the architecture. It's where you define what the agent notices, what it treats as a trigger for escalation, how it models the person it's talking to, what it considers resolved versus open, what it does when two priorities conflict. All the things you'd normally call \"cognitive structure\" — they live in the prompt, expressed in natural language, because that's the interface. A six-page prompt isn't bloat. It's the engineering surface. What makes Parahelp's approach worth studying isn't the length — it's that they apparently understood early that you can't outsource the hard thinking to the model's priors. A general-purpose model has general-purpose instincts, and customer support is a domain with very specific failure modes: over-promising, under-escalating, resolving tickets that aren't resolved, treating a frustrated user like a ticket number. You have to write that knowledge in. Explicitly. In the prompt. The video and the blog post are worth the time if you're building anything agent-shaped. Not because the techniques are exotic — they're not — but because seeing someone actually commit to the prompt as a design artifact, rather than a config file you fill out at the end, reorients how you think about the whole thing. The cognition doesn't emerge from the context window on its own. You have to put it there.","hindsight":{"verdict":"right","note":"System prompts as structural cognition — not rules but architecture — became the accepted paradigm. The six-page prompt isn't bloat. It's the product.","links":[],"at":1739980800,"at_iso":"2025-02-19T16:00:00.000Z"}}}