Every week someone shows me the "bad" output their AI wrote: a cold email that could have been sent to anyone, a discovery summary that restates the call transcript, an account plan that reads like a Wikipedia page. And every week the diagnosis is the same. They blame the model, or the prompt. Both are wrong.

The failure happens one step earlier. There was no brief.

The gap between facts and outreach

Most AI sales workflows look like this: paste in some account research, then ask for the email. The model gets a pile of facts and one instruction, "write me a cold email," and does exactly what you'd expect. It averages everything into something safe and generic, because you never told it what mattered.

The fix is embarrassingly simple. Split the job in two. First, turn the research into a short account brief. Then draft from the brief, not from the raw research.

The brief has exactly three lines:

  1. The trigger. What changed at this account recently. Funding, a new hire, an expansion, a bad quarter.
  2. The likely pain. What that trigger probably means for the person you're writing to, in their words.
  3. One specific fact worth referencing. Something that proves you actually looked. A quote from their earnings call, a detail from a job post, a number.

If you can't fill in all three, you don't have enough to write the email yet. That's the brief doing its job: it fails loudly instead of producing a polite nothing.

A prompt you can steal

Here is the brief-building prompt, word for word. Paste your research after it:

You are preparing a brief for a B2B sales rep. Read the account research below and return exactly three lines, nothing else. TRIGGER: the single most relevant recent change at this account (with its date). LIKELY PAIN: what this change probably means for a VP of Sales there, stated as a problem they feel this quarter. ONE FACT: one specific, verifiable detail from the research worth referencing in outreach, quoted with its source. Rules: if the research doesn't support a line, write "insufficient" for that line instead of guessing. Never invent dates, numbers, or quotes. Keep each line under 25 words. Research: [paste here]

Run this first. Read the three lines. Then, and only then, ask for the email, and paste the brief instead of the raw research. The difference in output quality is not subtle. Same model, same account, different input.

Why this works

The brief forces two things that raw research doesn't. First, it makes the AI commit to an interpretation (the pain) instead of hedging across all of them. Second, the "insufficient" rule stops hallucination at the exact point where most AI sales copy goes wrong: the model would rather invent a plausible trigger than admit it doesn't have one.

I've watched this pattern hold across hundreds of prompts. The prompts that perform are never the cleverest ones. They're the ones that name their required context up front, so you know before you run them whether you have what they need. Adoption follows the inputs, not the cleverness of the prompt.

Try it this week

Pick one account you're working right now. Run the brief prompt on whatever research you already have. If a line comes back "insufficient," that's useful information: go find that one fact before you write. If all three lines fill in, draft the email from the brief and compare it to what you'd have gotten from the raw paste.

Disclosure: I'm on the team behind Promptifi, a library of reviewed AI prompts and workflows for B2B sales. This brief-first structure is the same one behind our research prompts. You don't need our library to use the pattern above; it's yours. If you want 2,900+ more prompts built the same way, we're at promptifi.co.