System Prompt Generator

What belongs in a system prompt (and what doesn't)

A system prompt is the standing brief: who the assistant is, what it can do, how it talks, what it refuses. The user message is the request of the moment. Mixing the two is the most common reason a custom assistant drifts after ten turns.

Put in the system prompt anything you'd otherwise repeat:

• Role and audience. "You are a support agent for a bike shop, talking to customers who are not mechanics."
• Capabilities and boundaries. What data it has, what tools it can call, what it must never attempt.
• Tone. Short and blunt, or warm and thorough. Pick one.
• Output shape. Bullets? Max length? Always end with a next step?
• Standing rules: escalation, refusals, language, disclaimers.

Keep out of it: the specific question, today's data, one-off formatting requests. Those belong in the user turn. Paste a 4,000-word policy in because "it always needs it" and you've made every reply more expensive and slightly worse. Retrieve the relevant chunk per turn instead, and check the cost with the token counter.

There's a second reason for the split. System instructions carry more weight with most models and are harder to override mid-conversation. Anything you'd hate a user talking their way past belongs up top.

Order by importance, because the top gets read hardest

Models weight the start of the system prompt more heavily than the middle — the "lost in the middle" effect, reproduced across model families. Sequence deliberately: identity and mission, then hard constraints, then capabilities and tools, then tone, then output format, then examples if you need them.

If a rule genuinely cannot break, put it in the first five lines and say it once. Burying "never quote a price" in paragraph nine is how you end up quoting prices.

Write rules as instructions, not a wall of "don't"

"Don't be verbose. Don't use jargon. Don't hedge. Don't use emoji." Four negatives and the model still has to infer what you want. Positives give it a target instead of a minefield. Rewrite each prohibition as the behavior you want:

• "Don't be vague" becomes "give one concrete number or example per claim."
• "Don't write long answers" becomes "answer in under 120 words unless asked to expand."
• "Don't guess the plan" becomes "ask which plan they're on before answering billing questions."

Negatives still earn their place on hard safety lines — "never state a diagnosis" beats any positive phrasing. Use them where the boundary matters. Everywhere else, describe the target. Converting an old prompt full of prohibitions? Run it through the prompt rewriter, then check it with the analyzer.

Watch for rules that fight each other. "Be thorough and cite your reasoning" next to "keep answers under 50 words" produces a coin flip. Common prompting mistakes covers the usual collisions.

Out-of-scope requests and the ones it can't answer

Every assistant gets asked something outside its brief. If you don't say what to do, it improvises, and improvisation looks exactly like a confident wrong answer. Script three exits: outside scope, missing information, genuinely unknown.

"Never invent a slot" is a hard rule, so it stays a negative. "Ask one specific question and stop" replaces a vague "don't hallucinate."

Keep it short, then break it on purpose

A 2,000-word system prompt behaves worse than a 400-word one, and not just on cost. Rules dilute each other, contradictions creep in, and the model starts sampling from your instructions rather than following all of them. Aim for 150 to 500 words. Past that, you probably have three assistants pretending to be one.

Cut in this order: duplicated rules, examples that teach nothing the instruction already said, tone adjectives stacked four deep, and any rule you can't picture being violated. The optimizer spots dead weight fast.

Then test it before your users do:

• An empty or one-word message.
• A plausible but out-of-scope request.
• A user asking it to ignore its instructions.
• A question where the required information is missing.
• Something rude, and something in another language.

Run each five times. Models are sampled, so one clean pass proves nothing. When behavior wobbles, it's usually an ordering problem or two rules disagreeing. The system prompts guide walks through a full test pass, and the Claude prompt generator covers per-model format quirks. Needs multi-step reasoning? Pair the brief with a chain-of-thought structure instead of stuffing reasoning rules into the persona. The prompt library has briefs you can adapt.