A research assistant is not one clever prompt. It's a system: the same prompt every time, the same template every time, and a habit of checking sources before you believe a single number. That's what turns "I asked AI about this company" into research you can actually stand behind.
If you came from the video, that's the idea. Here's the part that doesn't fit in sixty seconds: the exact prompt I paste in, the template I dump answers into, a scorecard I score every company against, and the moment the whole thing quietly breaks if you're not watching for it.
Why a system beats a good prompt
The first time you research a company with AI, it feels like magic. The tenth time, you notice you asked slightly different questions each time, stored the answers in three different places, and have no way to say whether company A looks better than company B. You have ten piles of notes, not a comparison. The system fixes that. It makes every company go through the same funnel so the differences that show up are real differences, not just differences in how you happened to ask.
The four parts
An assistant that holds up has four parts. Skip one and it either drifts, forgets, or quietly hands you a made-up figure you repeat to someone who knows better.
One saved prompt
Same questions, same order, every company. The prompt below forces the AI to tag how solid each claim is: fact, their claim, or estimate. That tag is the whole game. An untagged AI answer blends a filing you could look up with a number it invented, and reads the same either way.
A fixed template
One document or table per company, always in the same shape: what it does, who pays, the numbers with sources and dates, your read, open questions. Sameness is what lets you line two companies up side by side later. If every note has a different structure, you can't compare, you can only re-read.
A source check
Treat the first draft as a list of leads, not answers. Open the sources it cites, confirm the two or three numbers that would actually change your mind, and downgrade anything you can't verify to "estimate". This is the step that separates research from vibes, and it's the one people skip when they're in a hurry.
A date stamp
Write the date you researched each company at the top of the note. Funding, pricing, and legal status change fast, so a six-month-old note is a lead, not a fact. When you come back, you'll know at a glance whether to trust it or refresh it.
The prompt I paste in
Same prompt, every company. Swap in the name, run it, and you get answers you can compare because they came from the same questions:
Research [COMPANY] like an investor. For EVERY claim, tag it:
[FACT] (with source + date), [COMPANY CLAIM] (from them, unverified),
or [ESTIMATE] (reported, not confirmed). If unknown, say "unknown".
Do not fill gaps with plausible guesses. "Unknown" is a good answer.
Cover, in this order:
1. What it sells, and the one-line pitch in plain language.
2. Who pays, what they pay, and how pricing is structured.
3. How it reaches customers (sales, self-serve, partners).
4. What makes it hard to copy, and what could stop it.
5. Which AI model or vendor it depends on, and the risk if
that vendor changes price, access, or terms.
6. Rough margins and any legal or regulatory exposure.
End with two lists: "What I could not find" and
"Numbers to verify before trusting."The "do not fill gaps" line matters more than it looks. Left to itself, a model will hand you a confident revenue figure for a private company that has never published one. Telling it "unknown" is allowed is how you keep that out of your notes.
A worked example: diligence on a pitch deck
Say you're an angel investor. A founder has sent a deck and wants a first meeting Thursday. You have maybe forty minutes and a deck that, like every deck, is written to make the company look inevitable. The assistant's job here is not to decide for you. It's to turn the deck into a list of things to confirm and a list of questions worth the founder's time.
You paste the deck text into the prompt above with the company name. The AI comes back with the pitch tagged as a company claim (correct, it came from them), the market size tagged as an estimate (also correct, decks inflate this constantly), and a "could not find" list that includes real revenue, churn, and who else is on the cap table. That last list is the useful part. It's the difference between walking into the meeting impressed and walking in prepared.
Then you do the source check. The deck says the founder previously "led growth" at a well-known company. You open LinkedIn, and the title was a mid-level role, not a leadership one. Small thing, but now you know the deck rounds up, so you read the rest of it a notch more skeptically. You drop each finding into the template and score the company on the card below. Forty minutes in, you have a one-page read and five questions no generic "tell me about your traction" opener would surface.
The scorecard
Once the notes are in the template, score the company so a gut feeling becomes something you can compare next week. Fill in one line per row: the AI's answer, then your confidence after the source check.
Company: __________ Researched on: __________
Answer Confidence (H/M/L)
Problem is real ______________ ___
Who pays, and why now ______________ ___
Real revenue (not ARR ______________ ___
run-rate on one deal)
Margins make sense ______________ ___
Hard to copy ______________ ___
Vendor / model risk ______________ ___
Founder claims check out ______________ ___
Top 3 things I could not verify:
1. __________
2. __________
3. __________
Questions for the first meeting:
1. __________
2. __________
3. __________What goes wrong, and what to watch for
The confident hallucination is the one that costs you. The AI states a revenue figure for a private company with no hedge, and it reads exactly like the facts around it. The fix is the tag and the source check, together. If a number is tagged as fact but has no source you can open, it's not a fact yet. Treat any unsourced number as an estimate until you've seen where it came from.
The other quiet failure is drift. You tweak the prompt for one interesting company, then keep the tweak, and three companies later you're no longer comparing like with like. Keep the prompt saved somewhere and paste the same one every time. If you genuinely need to change it, change it for the whole batch and note that you did, so you know your older scorecards used a different set of questions.
Staleness is the failure you won't notice until it embarrasses you. You cite a valuation in a meeting that was true two rounds ago. The date stamp is the whole defense: if the note is older than a few months on anything fast-moving, refresh it before you repeat it out loud.
And a scope trap worth naming: the assistant is great at what a company says about itself and slower on what others say about it. The deck, the website, and the press release are easy for it to summarize. The unhappy customer on a forum, the lawsuit, the quiet competitor eating the low end, those you often have to go find. Ask it directly for the bear case and the strongest criticism, and don't mistake a clean summary for a clean company.
This assistant is the front end of a bigger process. For the full framework behind these questions, see how to research a business, and when the company is specifically an AI startup, what I'd want to know before investing in an AI startup covers the vendor and model risk in more depth.