Artificial Lexicon
LLM guide

System prompts, plainly

The system prompt is the note taped to the model's desk before you walk in. Here is what to write on it.

Free5 min read

What a system prompt is

Every chat has two kinds of text. Your messages are the task. The system prompt is the standing instructions that apply to every task. In ChatGPT it hides behind "custom instructions". In Claude it sits in project instructions. In the API it is the system role.

The model reads it first and treats it with more weight than a normal message. That weight is the whole point. Rules that live there survive a long conversation better than rules you typed twenty messages ago.

What belongs in one

Four things earn a place. Everything else is rent.

  • Role. Who the model is for you. One line.
  • Audience. Who reads the output, and what they know.
  • Format rules. Length, structure, and what to leave out.
  • Boundaries. What the model should refuse to guess at, and ask instead.

The task never belongs there. If you write "help me plan a garden" in the system prompt, every future chat becomes about gardens. Standing instructions stand. Tasks pass through.

Write rules you can test

"Be concise" is a wish. The model nods and rambles. "Answers under 150 words unless I ask for more" is a rule. You can check it, and so can the model.

wish vs ruleWish: Be helpful and concise. Use good formatting. Rule: Answers under 150 words unless asked. Use a list only when there are 3 or more items. Never open with "Great question".

Every rule you can test earns its keep. Every rule you cannot test trains the model to skim.

Few strong rules beat many weak ones

Ten rules hold. Forty rules blur into a mood. When the list grows, group it: one block for voice, one for format, one for boundaries. Cut anything the model already does without being told.

💡 A short test: delete a rule. If the output does not change, the rule was decoration. Leave it deleted.

When the model drifts

Long chats erode instructions. The model follows the last few messages more than the first. Two fixes, in order:

  1. Repeat the one rule that broke, in your next message. Not the whole list.
  2. If drift keeps happening, the system prompt is too long. Shorten it until the survivors hold.

Two working examples

A writing assistant, in the voice this site uses:

writing deskYou edit my drafts. Audience: general readers, no jargon. Rules: - Short sentences. Plain words. No em dashes. - Cut adverbs unless the sentence dies without them. - Show the edit, then one line on why. - If a claim needs a source I did not give, flag it. Do not invent one.

And an art prompt assistant, built for the walls on this site:

prompt deskYou write Midjourney prompts from my rough ideas. Structure: [subject doing something], [one style word], [one lighting word], [2-3 details] --ar 3:2 Rules: - Under 40 words per prompt. - Concrete nouns only. Never "beautiful", "epic", "detailed", "8k". - Always give 3 versions that differ in exactly one word. - If my idea has no subject, ask for one. Do not pick for me.

Paste the second one into custom instructions and the image prompt workflow runs without setup in every new chat.

Published here first.

Keep going