Perplexity System Prompts: What You Can Customize

Perplexity has two separate layers where instructions live: the product interface available at perplexity.ai, which offers limited user-facing customization, and the Perplexity API, which accepts a full system message. Understanding which layer you are working in determines what you can actually change. Tools like fenbs keep the prompts and research rules you want to carry from one session to the next, outside the Perplexity interface itself.

The hidden product system prompt that Perplexity uses internally is not editable. What you control is a narrower set of preferences in the product and the full message structure in the API.

Product instructions vs. API system prompts

In the Perplexity web and mobile product, the main customization option is the AI Profile: a short description of who you are and how you want Perplexity to respond. This is not a traditional system prompt; it is a set of preferences Perplexity injects into its own system context. You do not see or edit the full prompt the product uses.

Through the Perplexity API (check the current endpoint and available models on the Perplexity developer documentation), you send a messages array with a system role message, just as you would with OpenAI or Anthropic models. This gives you full control over what the model receives as its instruction context. The API and the product are separate: a system prompt you set through the API does not affect the product interface.

What you can customize in the Perplexity product

In the Perplexity web interface, the AI Profile lets you describe your background, preferred response length, and subject areas you want Perplexity to prioritize. These preferences influence tone and focus but do not override Perplexity's core behavior: it still runs web searches, still cites sources, and still applies its own content policies.

• Background and context: who you are, your field, your experience level.
• Response style preferences: how detailed, how formal, how technical.
• Subject focus: areas you frequently research.

These settings apply across your Perplexity sessions on that account. They are not scoped to a project or a single query. If you need different behavior for different research tasks, the API is the more flexible path.

Setting a system prompt through the Perplexity API

The Perplexity API accepts requests in the same messages format as OpenAI's Chat Completions endpoint. You can add a system role message at the start of the array to set context for the model. This is useful for building research pipelines or applications where you need consistent behavior across many queries.

{
  "model": "sonar-pro",
  "messages": [
    {
      "role": "system",
      "content": "You are a research assistant for a US law firm. Summarize findings in plain language. Cite primary sources only. Do not state legal conclusions."
    },
    {
      "role": "user",
      "content": "What are the current FTC guidelines on endorsements?"
    }
  ]
}

One distinction worth knowing: Perplexity models run a web search as part of answering most queries. Your system prompt shapes how the model interprets and presents the search results, not whether it searches. If your use case requires suppressing the search layer entirely, check whether the model you are calling supports that option in the current API documentation.

For a comparison with context-building approaches that go beyond a single system message, context engineering vs. prompt engineering explains the broader picture.

What you cannot change

Perplexity's internal product system prompt is not exposed. You cannot edit the instructions that govern how its answer engine selects and presents sources, how it formats its default responses, or how it applies its content policies. Any customization you add sits alongside -- not instead of -- those hidden instructions.

This is similar to every hosted AI product: the vendor sets the base instructions and you add to them. Understanding this boundary prevents frustration when an instruction you set does not produce the change you expected. The model is following both your instruction and the product's, and the product's instructions usually take priority in a conflict.

Storing prompts worth reusing

A research prompt that works well for one project -- defining scope, source constraints, format requirements -- is worth keeping for the next one. Rewriting from memory produces a weaker version each time.

Testing your Perplexity API prompt

After writing a system prompt for the Perplexity API, send a few varied requests and check that the output matches your intent. Perplexity models run a web search as part of answering, so the test should cover questions that would normally return a search-heavy response. Verify that your format instructions, scope restrictions, and citation rules all hold.

Also test what the model does when the search results are sparse or contradictory. A well-written system prompt should tell the model what to say in those cases rather than leaving it to improvise. For example, if your prompt says "cite primary sources only," test with a question whose answer is only available in secondary sources and see how the model handles it.

How to do this with fenbs

1. Create a free account. Start with a personal kanban board for your research projects.
2. Add a task for each research prompt you want to formalize. In the task, paste the current version and note what it controls.
3. Once you test and approve a prompt, save it as a Skill in your fenbs project. Note the model it was tested on.
4. Write your standing research rules as Decisions and rules in fenbs -- for example, "use primary sources only" or "do not state legal conclusions." Connected assistants read these at the start of a session.
5. When a prompt needs updating, record why in the task and update the Skill so the next session uses the current version.

Related: perplexity MCP covers how to connect Perplexity to an MCP-enabled workflow. System prompt examples has patterns worth adapting for research contexts.