How to Write a System Prompt — Step by Step (With Examples)

A system prompt is the most leveraged piece of writing in any AI app. Get it right and the model behaves the way you want for every conversation. Get it wrong and you spend weeks chasing edge-case bugs in user-prompt logic that should have been fixed in two lines of system prompt.

This guide walks through writing a system prompt from scratch using a repeatable 5-step process. By the end, you'll have a production-ready prompt for your specific use case, plus a free generator that does steps 1-4 automatically.

Step 1 — Define identity

The first line establishes who the assistant is. This sets the tone for everything that follows. Models interpret all subsequent rules through the lens of this identity.

Bad: "You are an AI assistant."

Good: "You are a senior tax accountant assistant for US small businesses with revenues under $5M."

The good version sets a domain (US tax), a level of expertise (senior accountant), and a target audience (small businesses under $5M revenue). Each of those informs the model's word choice, depth of explanation, and assumed reader sophistication.

Step 2 — List capabilities

Tell the model what it can help with. Be specific. Models default to "I can help with anything" if you don't constrain them.

Bad: "You can answer questions."

Good:

You help with:
- Calculating quarterly estimated tax payments
- Explaining IRS forms (Schedule C, 1099, W-9, etc.)
- Identifying common deductions for sole proprietors and LLCs
- Distinguishing between business and personal expenses
- Bookkeeping basics for business owners

This list both tells the model what to focus on and acts as an implicit "do not help with" list — anything not on the list should trigger a redirect.

Step 3 — Write rules (always do)

These are the behaviors you want every response to exhibit. Pick 8-12 — past that, models start forgetting some.

Always:
- Confirm whether the user is a sole proprietor, LLC, S-corp, or C-corp before answering tax questions
- Cite the IRS publication or section when explaining a rule
- Distinguish between federal and state tax considerations
- Recommend consulting a CPA for situations involving more than $50K
- End each response with an offer to clarify or go deeper

Notice how each rule is concrete and observable. "Be helpful" can't be checked. "Cite the IRS publication" can.

Step 4 — Write constraints (never do)

These are the behaviors you want the model to always avoid. Often the most important section, because models will do unexpected things by default.

Never:
- Give specific dollar advice about tax strategies — always refer to a CPA
- Predict future tax law changes
- Discuss state taxes for states you don't have data on (only federal + CA, NY, TX, FL)
- Use legalese — always plain English
- Promise outcomes ("you will save $X")

Constraints are where you defend against the riskiest model behaviors. For a tax assistant, the risks are giving advice that gets the user in trouble. For a customer chatbot, they're promising things the company can't deliver. For a coder, they're using deprecated APIs.

Step 5 — Specify output format

Tell the model how to structure its responses. Without this, the model picks its own format and varies between responses.

Format every response as:
1. One-sentence summary of the question
2. The answer in 2-4 short paragraphs
3. A bulleted list of practical next steps
4. A "Related questions" section with 2-3 follow-up questions the user might have

Use plain English, not legalese. Bold key terms on first use.

Format rules are some of the highest-impact instructions you can give. They make every response feel uniform and professional, which is half the user experience.

The complete example

Putting all five steps together for the tax assistant:

You are a senior tax accountant assistant for US small businesses with revenues under $5M.

You help with:
- Calculating quarterly estimated tax payments
- Explaining IRS forms (Schedule C, 1099, W-9, etc.)
- Identifying common deductions for sole proprietors and LLCs
- Distinguishing between business and personal expenses
- Bookkeeping basics for business owners

Always:
- Confirm whether the user is a sole proprietor, LLC, S-corp, or C-corp before answering tax questions
- Cite the IRS publication or section when explaining a rule
- Distinguish between federal and state tax considerations
- Recommend consulting a CPA for situations involving more than $50K
- End each response with an offer to clarify or go deeper

Never:
- Give specific dollar advice about tax strategies — always refer to a CPA
- Predict future tax law changes
- Discuss state taxes for states you don't have data on (only federal + CA, NY, TX, FL)
- Use legalese — always plain English
- Promise outcomes ("you will save $X")

Format every response as:
1. One-sentence summary of the question
2. The answer in 2-4 short paragraphs
3. A bulleted list of practical next steps
4. A "Related questions" section with 2-3 follow-up questions

Use plain English. Bold key terms on first use.

That's a complete, production-ready system prompt. It's about 250 tokens, which is well within reasonable bounds for any LLM. You can paste it directly into the OpenAI API, the Anthropic API, the Gemini API, a custom GPT, or any other endpoint.

Iteration after the first draft

Your first draft is rarely your last. After deploying:

1. Test with edge cases. Off-topic questions, adversarial inputs, ambiguous requests, multi-turn conversations.
2. Note failures. When the model breaks character or violates a rule, write down what happened.
3. Refine constraints. Most failures are missing or vague constraints. Add a more specific "never" rule.
4. Re-test. Run the same edge cases again to verify the fix.
5. Avoid bloat. If a new rule makes you want to add 3 more, you may be over-fitting to one edge case. Step back.

Using the generator instead

The free system prompt generator automates steps 1-4 by giving you a use-case picker, a rules toggle list, and a custom instructions field. You spend 2 minutes selecting the right options instead of writing from scratch. Step 5 (output format) is still up to you to add or accept defaults.

The generator produces a prompt that's 80% of the way to production. You add the last 20% (specific constraints for your business, custom output format, few-shot examples) and you're done.