Perplexity Custom Instructions: Set Up Research Rules

Set Perplexity custom instructions in the Personalization section of your account settings; use Project instructions for rules that belong to one research effort. Write a short set of observable requirements, then test them with a new question before relying on them. Tools like fenbs can keep the approved prompt outside Perplexity so you can reuse it with other connected assistants.

The useful instruction tells you what to check: which sources count, how uncertainty appears, and what the answer should contain. A request to sound knowledgeable gives you no such test. Start with a real research task that you understand well enough to review.

How do you add custom instructions in Perplexity?

Perplexity’s Account Settings guide↗ (opens in a new tab) documents a free-text Custom instructions field under Personalization, alongside response length and formatting controls. Use the web settings for this setup:

1. Open your account settings through your avatar and All Settings, or go to perplexity.ai/account/details.
2. Choose Personalization and find Custom instructions. Add preferences you want to reuse, such as American spelling and a nontechnical audience.
3. Check Response Length and Headers and Lists so those choices match your written preferences. Avoid requesting a long report while selecting a short default.
4. Save your changes using the control shown in your settings, then ask a new research question and inspect the answer.

Keep an unchanged copy before you edit. If a new rule makes answers less useful, you can restore the earlier wording and compare the same question again. Do not add every preference at once: a small first version makes failures easier to explain.

Which rules belong at account level or in a Project?

Separate preferences about your answers from requirements for one assignment. Otherwise, an instruction written for a vendor comparison can become an awkward default for unrelated questions. This is the same practical distinction used when writing ChatGPT custom instructions.

For project-specific rules, select Projects in the sidebar, create or open a Project, then open Settings and Context. Some accounts instead show Configure project, See all settings, and General. Perplexity’s Project documentation↗ (opens in a new tab) places instructions there and says they are used for every query in the Project. Older guidance may call these workspaces Spaces.

Keep the approved source policy in fenbs’s Decisions and rules when other assistants must follow the same requirement.

What should reusable research instructions say?

Use five ingredients: audience, evidence, uncertainty, output, and boundaries. Tell the assistant what to do when evidence is missing instead of asking it to be confident. The following is a starting point for a small US design studio researching software; adapt its subject and deliverable to your work.

Audience: a small US design studio with no dedicated IT staff.
Use American English and explain unfamiliar technical terms.
For product features, prefer the vendor’s own documentation.
Place a source beside each material factual claim.
Separate documented facts from your interpretation.
If evidence is missing or contradictory, say what remains unverified.
Start with a direct answer, then give evidence and practical caveats.
End with the next check a person should make before choosing.
Do not describe a recommendation as an approved decision.

The rule about approval matters when research becomes a proposal. “This option appears suitable” and “the studio chose this option” describe different states of work. A generated answer should not silently cross that boundary.

Avoid a long imaginary biography or instructions to agree with your preferred conclusion. Include only background that changes the answer. For more examples of converting vague preferences into checkable behavior, see Copilot instruction examples.

How do you test whether the instructions work?

Test outputs rather than asking whether Perplexity understands the rules. An assistant can repeat an instruction and still miss it during research. Choose a few questions whose evidence you can inspect, and keep the questions unchanged when comparing revisions.

• Ordinary case: ask about a documented product feature. Check that the source is the vendor’s page and supports the exact claim.
• Missing evidence: ask whether a product supports a requirement you cannot find documented. Check that “not verified” stays distinct from “not supported.”
• Conflicting evidence: provide two dated pages that disagree. Check whether the answer identifies the disagreement instead of combining them into one confident claim.
• Audience check: ask for a technical explanation. Check that unfamiliar terms are explained and the requested structure remains readable.

Record the question, instruction version, result, and failure you observed. Change one instruction, repeat the failed case, then rerun an ordinary case to check that the repair did not make every answer cumbersome. These are your own checks, not a guarantee of future accuracy.

Why might Perplexity give a result that breaks a rule?

First check the location: are you asking inside the Project whose instructions you edited? Then compare the current question with those instructions. “Give an exhaustive report” and “use a short answer” ask for different outcomes; remove the contradiction before adding stronger wording.

Next inspect the evidence itself. Asking for official documentation does not make an unavailable page accessible, and asking for citations does not establish that each cited passage supports the answer. Open the sources behind the claims you plan to reuse.

If unrelated personal context keeps appearing, review Settings and Memory. Perplexity’s Memory guide↗ (opens in a new tab) documents separate controls for memories and search history. Inspect those influences separately from the instruction text; changing a prompt is not the same operation as clearing stored memories.

Use instructions as guidance you can review. Perplexity’s account help↗ (opens in a new tab) explicitly notes technical limitations with custom instructions. If a short, consistent requirement still fails, preserve a minimal example for support rather than treating repeated demands as a fix.

How do you keep the approved instructions reusable?

Save the wording that passed your checks, its purpose, and the date you reviewed it. Keep a proposed revision clearly separate from the approved version. When a requirement changes, record who chose the change and why so another assistant can understand it.

fenbs’s Skills and prompts hold reusable instructions for supported connected assistants. You copy the relevant wording into Perplexity yourself; fenbs does not connect directly to Perplexity or synchronize its settings.

How to do this with fenbs

1. Create a free account and add the research project to your personal kanban board. It is free for one person, with no credit card.
2. Save the reusable instruction in Skills and prompts. Include its purpose, the approved wording, and an example question for checking it.
3. Record the person’s continuing research requirements in Decisions and rules. Keep proposed changes open until a person chooses them.
4. Copy the relevant prompt into Perplexity and test it there. Connect another supported assistant to fenbs through Integrations when you want it to read the same approved prompt and project rules.
5. Save any useful failure as a lesson, and save Where we left off before the next session. Include which instruction version you tested and what remains unchecked.

If you research alone entirely within Perplexity and need only answer preferences, its own settings may be sufficient. fenbs becomes useful when the approved instructions and resulting work need to reach other assistants. MCP is a standard way for an assistant to use outside tools; the Perplexity MCP guide covers giving an assistant Perplexity research tools.

For app-specific setup elsewhere, use the guides to Claude custom instructions and NotebookLM custom instructions. The AI prompts library collects related ways to write and maintain reusable work instructions.