Prospect research & outreach drafting

Every line in the email opens a page you can read.

Dossier reads what a person has actually said in public, works out whether there is a real reason to reach out, and writes the message — showing every page it used, and everything it threw away.

Real run — nothing staged 47 sources · 22 person-level

Pedro,

Your breakdown on LightconePod about AI's impact on company building hits different regarding operational efficiency. Reps burning half their week on manual prospect research kills leverage, so I built Dossier to deliver source-verified intelligence in ninety seconds.

Hook
First-party post, 10 June 2026 — person-level, chosen across 8 candidates
Source
x.com/ycombinator/status/2064728817802092579
Checked
50 of 50 facts verified against the text of the page they came from
Dropped
8 copies of the same story folded into one; 3 facts about a different person with the same name — nothing tied them to Brex

The problem

Doing it properly takes too long. Doing it with AI reads like AI.

Researching one prospect well is roughly forty minutes of work. Nobody has that per name — so the job goes to a tool, and two problems arrive with it.

Problem 01

It takes forty minutes

Per prospect, done properly. That time is the reason the research gets skipped and a template goes out instead.

Problem 02

The message reads robotic

Competent, polite, obviously generated. The recipient can tell, which costs you the reply it was meant to win.

Problem 03

It surfs, it doesn't read

Tools skim the open internet rather than opening the source a claim came from. So you can't tell where a line came from — or whether it's true at all.

Two ways in

It depends on one question: do you already have the names?

Same engine underneath either way. The only difference is where you start — with a person, or with a market.

Leads

You know who to talk to. You don't have a reason to open the email.

You have
A list — CRM export, event attendees, inbound signups.
You give
People.
You get
A researched, sourced message for each one.

Outbound

You have no list at all. You have one company that worked.

You have
A customer you just won, or a company you sell against.
You give
One company name.
You get
Their competitors, the right people inside them, each already researched.
1 Targeting 2 Identity 3 Research 4 Extract 5 Ground 6 Judge 7 Draft

Both routes run the same seven stages. A name that came out of a campaign is researched as carefully as one you typed in yourself — there is no weaker second path.

Feature 01

Not another web search tool.

It reads their LinkedIn and their X, not the company homepage

The easy build is a wrapper: hand a model a name and ask what's new. It answers from memory — confidently, with no date and nothing you can check. It will write you a funding round that never happened.

Dossier doesn't ask a model what it remembers. It goes and reads — their LinkedIn profile and posts, what they put on X, and the podcasts, interviews and talks they have actually spoken on. What is about the person is kept apart from what is about the company, and scored on its own — because company news is what everyone else in that inbox is already sending.

One real run, counted honestly
47sources readpages
22about the person, not the companyperson-level
50concrete facts extractedcandidates
11thrown out — 8 repeat copies, 3 about the wrong persondropped
1became the hook — with its source attachedused

And the check is dumb on purpose

Every fact has to bring exact words from the page it came from, and those words are checked back against that page before anything is written. No model is ever asked whether a fact is true.

A model that invented the fact will happily confirm it. A string comparison can't be talked round.

It will hand you an honest blank before it hands you a convincing lie.

When there's nothing real to say, it says so.

When a run finds nothing it can check, it says no signal found, and marks the draft generic instead of dressing it up.

If the search budget runs out, it says research failed — never “this person has nothing public”. One is about the person. The other is about my bill.

Feature 02

A voice that gets closer to yours every time you correct it.

You start at one, not zero

You design your persona on day one, so the first message already sounds like you rather than like a language model. Then it watches what you change.

Cut the closing line three times and it writes that rule down — and from the next draft on, the line is already gone. You never train it. You just keep editing, and it stops needing to be edited.

Because one edit is just a mood.

How a rule earns its place
+The same change, made againadds on
A draft sent unediteddowngrades
3consistent changes before a rule is written downthreshold

Research you can check, in a voice that's yours.

Two ways in depending on whether you have the names yet, every claim tracing back to a page you can open, and a persona that learns how you write by watching you correct it.