Inside Perplexity's Search Agent Prompt: The Minimal Rules That Make It Fast and Accurate

Overview

Perplexity is a search‑focused agent whose system prompt is roughly 1,000 words and defines eight tools: search_web, get_full_page_content, search_browser, open_page, close_browser_tabs, execute_code, search_people, and load_skill. The prompt consists of only two blocks — tool_policy (tool‑use rules) and response_style (answer‑style rules) — because the product does not need to teach the model how to converse, only how to search.

Tool‑Use Rules ( tool_policy )

1. Mandatory tool call before answering

Prompt fragment:

Use at least one relevant tool before answering unless tools are unavailable, disabled by a tool result, or unnecessary for the specific environment.

Analysis: This turns "search first" from a suggestion into a hard constraint, preventing the model from relying on stale training data. The three exemptions (unavailable, disabled by result, environment‑specific) are precisely worded to close loopholes. The author notes this is a zero‑cost reliability lever: one system‑prompt line changes default behavior from "generate" to "retrieve then answer."

2. Decompose complex queries into independent sub‑questions

Prompt fragment:

For complex queries, break the request into independent subquestions and gather information efficiently.

Analysis: Breaking a large query (e.g., "compare three companies' latest earnings") into independent sub‑queries lets each map to a precise search term, improving recall. "Independent" is key — it enables the next rule (parallel calls). The goal is retrieval efficiency, not output structuring.

3. Prefer parallel tool calls when independent

Prompt fragment: Prefer parallel tool calls when calls are independent. Analysis: Parallelizing independent lookups cuts latency roughly in half. The condition "when calls are independent" forces the model to judge dependencies; "Prefer" rather than "must" preserves a fallback to serial execution. The author calls this a "one‑line prompt for free latency optimization."

4. Hard limit of three tool calls per turn

Prompt fragment:

Make at most three tool calls before producing the final answer.

Analysis: The cap controls cost and latency without gutting coverage. Without a limit, agents tend to exhaust their allowance or loop. Three calls is a deliberate trade‑off: information not found in three tries likely indicates a query or source problem, and marginal returns diminish. This limit pairs with the next rule for a double safety net.

5. Stop‑condition self‑check after each call

Prompt fragment:

After each tool call, decide whether the available information fully answers the user's request. Continue only when more information is genuinely needed and the tool‑call limit has not been reached.

Analysis: Checking after every call (not after the budget is spent) lets simple queries stop early. "Genuinely needed" suppresses the tendency to over‑search. Continuation requires both conditions — information gap and remaining budget — creating a natural efficiency habit.

Answer‑Style Rules ( response_style )

1. Direct answer in the first one or two sentences

Prompt fragment:

Start with a direct one‑ or two‑sentence answer to the user's core question. Do not put a heading before the opening answer.

Analysis: Search users want quick confirmation; the first sentence delivers the core answer so users can leave after a glance. Banning even a heading before the answer shows a pixel‑level obsession with "first screen = answer." This improves both perceived and actual speed.

2. No nested bullet points

Prompt fragment: Keep bullets top‑level only. Do not nest bullets. Analysis: Nested lists disguise parallel facts as hierarchies, forcing readers to mentally track indentation. Search answers are typically flat facts, so a flat layout matches the content. The author states this single rule would improve a huge number of AI‑generated answers — a one‑token rule that eliminates a whole class of formatting accidents.

3. Avoid repetitive summaries unless explicitly requested

Prompt fragment:

Avoid repetitive summaries and conclusions unless the user explicitly asks for one.

Analysis: Models habitually append "In summary…" sections that waste tokens and user time. The exception (explicit user request) draws a clean boundary. Combined with the direct‑opening and no‑nesting rules, the trio creates the "faster, cleaner" feel users associate with Perplexity.

Core Takeaway

Vertical agents don't need long prompts — they need precise rules for when to use which tool and when to stop . Search agents are about tool discipline, not conversational skill: tool‑first, three‑call ceiling, direct answer without headings. Tool count matters less than each tool having a clear, narrow purpose. Perplexity's essence: it doesn't teach the model how to chat; it teaches the model how to search — and the more precise the rules, the more reliable the result.