How to Prompt an AI Research Agent: A Template That Actually Works

ON THIS PAGE · 7 SECTIONS

1. The six parts every research agent prompt needs
2. The template, written out
3. A weak prompt versus a working one
4. Does telling it to act like an expert make it more accurate?
5. What should the agent say when it does not know?
6. How to test a prompt before you trust it on the full list
7. Questions about prompting an AI research agent

An AI research agent prompt is the instruction set that tells a model which sources to read, what fields to fill in for each account and what to do when the answer is not in front of it. Most research agent prompts fail on the third part. They tell the model what to find and skip the rule for a missing answer, so the model fills the gap with a plausible guess instead of an honest "unknown." A prompt that works names the role, the exact sources, a fixed output schema and the rule for a missing answer, written out so the model has no room to improvise around any of the four.

The six parts every research agent prompt needs

Teams that get a usable brief out of a research agent on the first few tries are not writing cleverer sentences. They are including six parts that a vague prompt leaves out. The parts are ordered so the model reads its boundaries before it reads its task.

1. Role and boundaryOne sentence that states the job and rules out the jobs next to it: this agent writes a research brief, not outreach copy. It does not decide whether the account fits your ICP.
2. Inputs, named exactlyThe specific sources it may read: this CRM record, the company site, the two most recent job posts, named news from the last 90 days. Everything outside that list is off limits, including the model's own memory.
3. The reason this account came up nowThe signal or trigger that put the account on the list, so the model researches the right moment instead of writing a generic company summary a search engine could produce on its own.
4. A fixed output schemaA named list of fields, each with an allowed answer or "unknown," never a free text paragraph the model can pad with filler to look complete.
5. An evidence line per fieldThe exact sentence or URL each answer came from, so a reviewer checks the claim in seconds instead of re-researching the account from scratch.
6. A rule for disagreement and silenceWhat happens when two sources conflict and what the model writes when no source answers a field. Both cases need an instruction or the model picks one for you.

The template, written out

Here is the shape, with the lines you would actually put in front of a model. Keep the order. A model that reads its sources and its schema before it reads the account details stays inside both far more reliably than one that reads the account first and gets told the rules afterward.

Role. "You are preparing a one page account brief for a B2B sales rep. You do not write outreach copy and you do not judge whether this account fits our ICP. Your only job is to answer the fields below using only the sources given."

Inputs. "You may use only: the attached CRM record, the company's own website, the two most recent posts on its careers page and any news article from the last 90 days that names the company directly. Do not use anything from training memory. If a fact is not in one of these sources, treat it as not known."

Output schema. "Return exactly these fields and nothing else: company_summary (one sentence), likely_buyer (name and title or 'unknown'), trigger (the one recent event that applies with its date and source), tech_change (a named tool or 'unknown'), confidence (high, medium or low) and source (the exact sentence or URL behind every field above)."

Disagreement and silence rule. "If two sources disagree on a fact, report both and set confidence to low. If nothing in the sources answers a field, write 'unknown.' Do not infer an answer that sounds likely."

One worked example. Paste one filled out example row before the real batch runs, built from an account the reviewer already checked by hand. A model copies a pattern it can see far more reliably than a pattern it was only told about in a sentence of instructions.

A weak prompt versus a working one

A locked schema fixes the format. It does not fix the truth. Only named sources and a review step do that.

Does telling it to act like an expert make it more accurate?

Not on its own. The research on this is specific enough to act on. Giving a prompt a confident persona changes tone and formatting. It does not improve the facts the model pulls from memory or the sources it was given. On knowledge heavy questions it can pull accuracy down instead of up. The two controls that actually move accuracy are the ones in the template above: a named source list and a locked schema with an honest "unknown."

Source: OpenAI, Introducing Structured Outputs in the API, 2024, arXiv, Expert Personas Improve LLM Alignment but Damage Accuracy, 2026, Digital Applied, 2026

The persona study is worth sitting with because it cuts against a habit a lot of teams pick up early: dressing up the first line of a prompt with "You are a world class B2B research analyst with twenty years of experience," on the theory that a stronger identity produces a stronger answer. On tasks that test stored factual knowledge, researchers found the opposite. The persona shifts the model into what the paper calls an alignment mode, where it optimizes for sounding like the role instead of checking what it actually knows. Write the role line for clarity on the job, not as a motivational speech.

What should the agent say when it does not know?

"Unknown" is the single word doing the most work in a research agent prompt that holds up at volume. It is also the word most teams forget to ask for. A schema without that option still looks disciplined, five or six clean fields with one allowed shape each, but it still produces a guess the moment a source runs dry because the model has nowhere else to put the answer. We wrote about this same gap in how LLM lead enrichment avoids hallucinated fields: a model with no honest way to say "not found" will manufacture a plausible one instead. A fabricated title or funding round reads exactly like a correct one once it sits in a CRM field.

Building the schema is the easy half. The harder habit is treating a batch of briefs with a lower fill rate as a better batch, not a worse one, as long as the unknowns are honest. A rep who sees "unknown" on the likely buyer field knows to check LinkedIn by hand before sending anything. A rep who sees a confident name that was never actually sourced has no reason to check at all, until a reply comes back confused about who they are and why they were contacted.

How to test a prompt before you trust it on the full list

Run a new or changed prompt on 15 to 20 accounts you already know well enough to grade the output by eye, before it ever touches the 50 to 100 account pilot you will want for the wider rollout of a research agent. Read every field against the source it cites, not just the fields the model flagged as low confidence. The failures almost never spread evenly. They cluster in one field type: most often the buyer title because job titles on a careers page lag the org chart or the trigger field because a news search returns a story about a different company with a similar name.

Log each failure by field, not by account. After a dozen or so test runs, a pattern shows up: the buyer_title field keeps citing a careers page post from eight months ago or the trigger field keeps pulling a funding round that belongs to a sister company in the same portfolio. That pattern points at the exact line in the prompt to fix, usually the source list or the disagreement rule. It is a faster fix than reviewing twice as many accounts by hand ever would be.

A prompt that passes this test is not finished forever. New sources get added, account mixes shift and a schema built for enterprise accounts tends to return more unknowns than it should once you point it at ten person companies with no press coverage. Treat the prompt the way you treat the rest of the outbound system: something that gets reviewed on a cadence, not something you write once and leave running.

Questions about prompting an AI research agent

What is the most important line in an AI research agent prompt?

The rule for a missing answer. A named source list and a locked schema still produce a confident guess the moment a source runs dry unless the prompt explicitly allows the model to write "unknown" instead.

Does telling the model to act like an expert improve its research?

Not for factual accuracy. A 2026 study found a detailed expert persona pulled MMLU accuracy down from 71.6% to 66.3% against a plain prompt on the same questions. A persona changes tone and format, not what the model actually knows or can source.

How specific should the source list in the prompt be?

Name the sources, not a category. "The company's careers page and news from the last 90 days" beats "recent public information" because the second version leaves the model free to count a training-memory guess as a source.

How many accounts should I test a new prompt on first?

Fifteen to 20 accounts you already know well enough to check each field by eye. Fix what the failures show you before widening to a 50 to 100 account pilot. Only move to the full list once that pilot holds.

Can one prompt work for every account tier?

Usually not without adjustment. A schema built for well covered enterprise accounts returns far more honest "unknowns" once pointed at small companies with thin websites and no press. Expect to loosen the source list or add a fallback field for the lower tiers rather than forcing one template to fit both.

Sources: OpenAI, Introducing Structured Outputs in the API, 2024, arXiv, Expert Personas Improve LLM Alignment but Damage Accuracy, 2026, Digital Applied, 2026, Anthropic, Claude prompting best practices