How Do You Write a System Prompt for a Sales Research Agent?

How do you write a system prompt for a sales research agent?

Give the agent a clear role, the reason the research matters, your ideal customer profile, the exact questions to answer, the sources it may use, and a strict output format. Tell it to cite a source for every fact and to say "not found" instead of guessing. Then test it on accounts you already know well.

In my view, sales research agents are one of the most useful things a GTM team can build with AI right now. Before a rep calls an account, the agent reads the company's site, recent news, job posts, and your CRM notes, then writes a short brief. Done well, it saves reps real time. Done badly, it invents facts that embarrass them on the call.

The difference is mostly in the system prompt. This article is how I write one, section by section, and why each part is there.

What is a system prompt in a research agent?

The system prompt is the standing set of instructions the agent follows on every run. It defines who the agent is, what job it does, what it may and may not do, and what its output should look like. The account being researched changes each time. The system prompt stays the same.

Anthropic's prompting guide for Claude says that setting a role in the system prompt focuses Claude's behavior and tone for your use case, and that even a single sentence makes a difference. For a research agent, that role sentence sets the tone for everything that follows.

Think of the system prompt as the job description you would give a new research analyst on day one. If a smart person could not do the job from it, the agent will not either.

What role should the research agent have?

Give it a specific, modest role, such as "You are a B2B sales research analyst preparing a pre-call brief for an account executive." Specific roles focus the output. Grand roles, like "world-class sales strategist," invite opinions and speculation, which is exactly what you do not want in research.

The role should also name the reader. A brief for an account executive about to run a discovery call needs different detail from a brief for an SDR deciding whether to send a first email. Say who reads it and what they will do next.

I add one sentence about tone: plain, factual, short. Research briefs that read like marketing copy get skimmed and distrusted. Briefs that read like a careful colleague's notes get used.

Why should the prompt explain the reason behind the research?

Explaining why helps the agent choose what matters. Anthropic's guide notes that giving context or motivation behind your instructions can give Claude more information about your goals and help it deliver more targeted responses. A research agent that knows the goal is a first discovery call will prioritize differently than one told only to "research this company."

In practice, I write two or three sentences of context. What the company sells, who it sells to, and what the rep wants to learn before the call. For example, whether the account uses a competing tool, whether it is hiring for roles that signal the problem you solve, and who likely owns the budget.

This context is where your ideal customer profile belongs. Paste a short version of it into the prompt, so the agent can judge fit, not just collect facts.

Which questions should the agent answer?

List a small, fixed set of questions, usually five to eight, in priority order. Fixed questions make briefs comparable across accounts and easy for reps to scan. Open-ended instructions like "find anything useful" produce long, uneven briefs that bury the one fact the rep needed.

A typical set for a B2B SaaS team might be: what the company does in one sentence, approximate size and stage, signs of the problem you solve, tools they appear to use, recent changes such as funding or leadership hires, likely buyer roles, and any prior history in your CRM.

For each question, say what counts as a good answer. "Approximate size" should state a range and where it came from. "Signs of the problem" should quote the source, such as a job post line. Clear standards per question are what turn an agent from chatty to useful.

How do you stop the agent from inventing facts?

Require a source for every factual claim, allow "not found" as a valid answer, and tell the agent that a missing fact is better than a guessed one. Ask it to quote the relevant text before it summarizes. Anthropic's guide recommends asking Claude to quote relevant parts of documents first, to help it focus on what matters.

The "not found" rule is the most important line in the whole prompt. Without it, agents tend to fill gaps with plausible guesses, because the format seems to demand an answer. With it, the brief shows clearly where research came up empty, which is useful information for the rep too.

I also tell the agent to separate facts from inferences. A fact has a source. An inference is labeled as an inference, with the reasoning stated. That one distinction makes briefs much safer to use on a live call. The same discipline is behind building a claims ledger for AI-drafted content.

Which sources and tools should the agent be allowed to use?

Name the allowed sources explicitly: the company's own website, its job postings, its press releases, your CRM records, and any enrichment tool you connect, such as Clay or Apollo. Tell the agent which sources to trust most when they disagree. Forbid sources you do not want, like forums or unverified directories.

If the agent uses tools through MCP servers or an API, list each tool and when to use it. Agents work better when they know which tool fits which question. For example, use the CRM tool first to check history, then the web tool for public information.

Log every tool call the agent makes. When a brief contains a wrong fact, the log tells you whether the source was wrong or the agent misread it. I covered how to set that up in logging the tool calls an MCP agent makes.

What output format works best for reps?

Use a fixed, short structure with clear section labels, one or two sentences per answer, and a source next to each fact. Put the most decision-relevant item first, usually fit and the main signal. Anthropic's guide suggests using XML tags to separate instructions, context, examples, and inputs, which also helps keep the output format consistent.

Show the agent an example of a good brief inside the prompt. The same guide says examples are one of the most reliable ways to steer output format and tone, and suggests three to five well-crafted examples for best results. Even one strong example helps a lot. Mark examples clearly so the agent does not copy their facts.

Keep the whole brief readable in under a minute. Reps read these between calls. If the brief needs scrolling, it is too long, and the important line will be missed.

How do you test the system prompt before reps use it?

Run the agent on ten accounts you already know well, including a few where public information is thin. Check every fact against its source. Count invented facts, missing sources, and answers that should have been "not found." Fix the prompt and rerun until the invented fact count is zero on your test set.

Include tricky cases on purpose: a company with a common name, a recent rebrand, a subsidiary of a larger group. These are where research agents most often mix up entities. If the agent confuses two companies in testing, it will do the same in production.

Write the prompt as a versioned spec, with a date and a change note each time you edit it. The format I use for other AI steps is described in writing a prompt spec for an AI step in a Zap, and it works just as well for agents.

What should you do next?

Draft your research agent's system prompt this week using six parts: role, context and ICP, fixed questions, sourcing rules with "not found" allowed, allowed tools, and output format with one example. Test it on ten known accounts. Ship it to reps only when every fact in the test briefs traces to a source.

My view is that the value of a sales research agent is trust, not speed. A brief that is fast but sometimes wrong teaches reps to ignore it. A brief that is careful, sourced, and honest about gaps becomes part of how the team sells.

If you want help building a sales research agent or the GTM systems that feed it, reach out through pravinkumar.co. I build websites and the automations behind them for lead generation, and research that reps actually trust is a big part of that work.