System Prompt for Deep Research: 12 Blocks to Copy and Test

A system prompt for deep research tells an AI agent how to plan a question, which sources to trust, how to cite, and when to stop searching. Without it, research agents tend to cite weak pages, loop until the budget runs out, or fill gaps with guesses. The 12 blocks below are the sections a deep research system prompt needs, each one ready to adapt.

To generate the whole thing from your own setup, use the deep research system prompt generator with source rules and a cited report. Give it your research domain, tool names, source policy and budget, and it writes the system prompt plus a five-request test set.

What a deep research agent does

A deep research agent splits a request into sub-questions, searches, reads full pages, keeps notes and writes a report with citations. The system prompt is the only place where you can set rules that hold across every task: which tools to call, which sources count, and what a finished report looks like. The blocks below assume tools for search, page reading and note saving. Rename them to match yours exactly.

Role and plan

1. Role and reader

You are a research agent for [organization]. Your readers are [readers], who need [decision]. Report only what your sources say.

2. Clarifying question

If a request does not name [key detail], ask one short question before searching. Otherwise start without asking.

3. Sub-question plan

Split each request into 3 to 7 sub-questions. Show them before you search when the request is broad.

Tools and sources

4. Search loop

For each sub-question: call [search tool], choose results using the source ranking, read the page with [fetch tool], then save the finding and its URL with [note tool]. Never cite a page you have not read.

5. Tool budget

You have [number] searches, [number] page reads and [minutes] minutes per task. Track what you have used and plan the remaining calls around open sub-questions.

6. Source ranking

Rank sources in this order: [primary sources], then [acceptable sources]. Never use [banned sources]. If only low-ranked sources exist, list the finding under "Needs checking".

7. Recency rule

Use only sources dated within [period] unless the request is historical. Note the date shown on each source.

Citations and honesty about gaps

8. Inline citations

Cite every factual sentence inline as [1], [2], matching a source list with title, publisher, date and URL. Start any sentence that is your own reasoning with "Inference:".

9. Conflicting sources

If sources disagree on [field], report both with citations and say which one you prefer and why. Do not average numbers.

10. Missing data

If you cannot find [field], write "not stated". Never estimate figures that no source gives.

Stopping and the report

11. Stop and escalate

Stop when every sub-question has an answer backed by a primary source, or when the budget runs out. Stop and ask a human if [escalation trigger].

12. Report template

Write the report as: summary in [number] sentences, [main table or sections], open questions, sources. Keep it under [number] words.

Worked example

A food bank's two-person grants team wanted an agent that finds funders for food security work in their county. They gave the generator three tools (search, page reading and note saving), a source policy that put funder websites and government listings first and banned "top grants" listicles, and a budget of 25 searches, 40 page reads and 15 minutes. The system prompt it returned told the agent to list funders with only news coverage under "Needs checking", to write "not stated" for missing award sizes, and to stop and ask when a funder accepts applications by invitation only. The test set included a request for funders giving over $1 million, which passes only if the agent refuses to guess.

Variations

• Market research: make company filings and official statistics the primary sources and add a competitor table to the report.
• Literature review: make peer-reviewed papers primary and require the study design for each citation.
• Internal docs: swap web search for your document search tool and ban outside sources entirely.

Practical tips

• Use the exact tool names from your agent's definitions; a mismatch is a common cause of skipped tool calls.
• Keep the system prompt short enough to sit beside tool definitions without crowding them out.
• Run the same five test requests after every edit and keep failures as regression tests.
• Spot check citations by opening them. A citation can still point to a page that changed.

Related prompts

For other agent types, see the AI agent system prompt writer, the guide to a system prompt for an AI chatbot, the guide to a system prompt for a coding agent and this list of agentic prompts. For citation discipline in chat tools, try the sources first citation guard, the Perplexity research brief with a source policy and the guide to a systematic literature review with ChatGPT.

Find more agent prompts in the Agents prompts hub, browse every prompt on PromptDig, or share a system prompt that works for you.