How to prompt Perplexity Sonar

Prompt Perplexity by splitting standing rules from the question: instructions hold role, tone, and grounding, and the input is the actual query that also seeds search. Perplexity’s prompt guide describes an agent loop that can call tools such as web search, read the result, and then answer. The same text therefore shapes the search and the final write-up. A vague input produces a vague search. A specific input, with the subject and the constraint named, is what that guide says improves retrieval.

Put rules that must survive every turn in instructions: how to cite, what to do when nothing relevant comes back, and the format of the answer. The guide says those instructions are re-read on every turn, so a long block costs more than it does in a single-shot prompt. Hard filters — date range, region, allowed domains, maximum steps — belong in request parameters, not in a paragraph the model may treat as a suggestion.

If search still returns nothing relevant, the documented behavior to ask for is an explicit admission, not a nearby substitute. Their example cites by domain rather than pasting full URLs.

The prompt structure

Task:
Compare Postgres logical replication and Debezium for streaming
changes into Kafka.

Context:
Postgres 16 on AWS RDS. Team of three, no dedicated data engineer.
Timeframe: sources from 2025 onward.

Constraints:
- Compare only on: setup effort, operational overhead, schema changes.
- Cite a source for every non-obvious claim.
- If sources conflict, show both positions.
- If a claim can't be supported, say so. Never invent links.

Output format:
Comparison table, then "Sources" as a numbered list.

The citation policy is not optional

Ask a search model for “sources” without forbidding fabrication, and you may get neat-looking URLs that don’t exist. Spell out the rules:

• Cite sources for figures, dates, product claims and quotes.
• If sources disagree, report both instead of picking one silently.
• If nothing supports a claim, say it couldn’t be verified.
• Never invent paper titles, DOIs, news links or social posts.

Don’t put example URLs or dummy citations in the prompt. They become templates for invented ones.

Separate facts from interpretation

Ask for Findings (or Answer), then Sources. If you want analysis, request a clearly labelled Interpretation section, so readers can tell verified information from opinion.

Scope the search

Underspecified questions pull in generic web filler. Add a scope sentence: date range, region, product version, or the specific criteria to compare. Don’t request an open-ended literature review unless that’s actually the goal.

Keep your beliefs out of the facts

If you already suspect an answer, label it as a hypothesis in Context (“our assumption is…”) rather than stating it as fact. Sonar should test it, not repeat it back.

Don’t ask it to browse a specific URL unless you give that URL.

Success criteria for research

For research that matters, check citations, not writing flair: every statistic maps to a source, and conflicts between sources are named.

Why Sonar prompts go wrong

• Chat-style questions that never mention sources.
• Asking for links without forbidding invented ones.
• Personal opinions in Context presented as verified facts.
• Example bibliographies the model copies.
• No date or version scope, so results are generic or stale.

When does a Perplexity prompt search the wrong thing?

The question is the search query, and it is vague. The prompt guide says input seeds the first search. “Write about our pricing” does not name the product, the old price, or the new price, so retrieval has nothing specific to fetch. Put the concrete terms in the question even if they also appear in instructions.

Standing rules are pasted into the question. Role, citation format, and “say when you find nothing” belong in instructions, which apply on every turn. Stuffing them into the query makes the search text longer than the subject. The guide also says to keep instructions short because they are re-read each turn.

A date range written only as prose. “Recent sources” is not a filter. The guide tells you to put retrieval limits such as date range, region, and allowed domains in request parameters. A sentence in the prompt is not that parameter.

Silence treated as permission to substitute. The documented grounding rule is to say when searches return nothing relevant, including after a rephrase, rather than swapping in an off-topic match. A prompt that only says “cite sources” never asks for that admission.

Keep findings and interpretation in different sections when you want both. A labeled Findings block and a labeled Interpretation block are how a reader can tell a retrieved claim from your gloss. Do not plant an example URL in the prompt. The guide’s citation pattern is a domain in parentheses, and an example link is a template an answer can imitate. If you already believe the answer, mark that belief as an assumption in the question rather than as a fact the search is supposed to confirm.

What does the same rough prompt look like for Perplexity?

Same draft as the other guides. The instructions block is the policy. The input is the searchable question. No extra product facts are introduced.

write a blog post announcing our windows app is moving from 1000 one-off credits to a flat 13.99 a month plan, about 600 words, its for people already using it, dont be hypey

Instructions:
You are drafting an announcement for people who already use the product.
Cite sources by domain for any claim that is not in the question.
If search returns nothing relevant, say so. Do not invent URLs.
About 600 words. No hype adjectives.

Input:
Write the announcement that this Windows app is moving from 1,000 one-off
credits to a flat 13.99 a month. Readers already use the app. Name both prices.