How to Prompt AI to Write a Founder Bio

Asking a model to write your bio produces the same paragraph every time: a visionary leader passionate about innovation, with a track record of driving growth. It is not a prompting failure exactly. If you prompt AI to write a founder bio without giving it any facts about you, filler is the only honest output available, and the model fills the gap with the average of every bio it has ever read.

The fix is to spend five minutes on a fact sheet before you write a single instruction. Everything else in this guide depends on that.

Step 1: the fact sheet

Open a note and write numbered lines. No prose, no adjectives, just checkable facts. Number them, because the verification pass later depends on it.

1. Role: co-founder and CTO, Merrow Labs, since March 2022
2. Before: six years backend engineer at Castellan (payments), led the team of 4 that shipped their card tokenisation service
3. Built: open-source library `ledgerfmt`, about 2,400 GitHub stars
4. Education: BSc Computer Science, University of Leeds, 2014
5. Public: spoke at PyCon UK 2024 on idempotency in payment systems
6. Current focus: reconciliation tooling for small finance teams
7. Based: Manchester, UK
8. Personal detail I am willing to use: restores mechanical typewriters

Two rules for this list. Only facts you would be comfortable seeing quoted back at you in a hostile room. And no claims you cannot evidence, because the model will happily promote a thing you did once into a thing you are known for.

If you are stuck on what belongs here, the test is whether a stranger could verify the line. Led the team of four that shipped tokenisation is verifiable. Passionate about fintech is not, and it is exactly the kind of sentence you are trying to avoid.

Step 2: the prompt

Now the instruction. The structure that works gives the model the facts, the audience, a banned list, and three length targets in one pass, so you get a family of bios that agree with each other rather than three separate attempts that contradict.

Here is my fact sheet. Each line is numbered.

<facts>
[paste the numbered list]
</facts>

Write three versions of my professional bio from these facts only:
A) 25 words, third person, for a conference speaker listing
B) 60 words, third person, for the about page of our company site
C) 120 words, first person, for a funding application

Rules:
- Use only what is in <facts>. If something is missing, leave it out.
  Do not infer seniority, scope, outcomes or motivation.
- No adjectives about me. Describe what I did, not what kind of
  person that makes me.
- Banned words: visionary, passionate, innovative, thought leader,
  guru, seasoned, dynamic, results-driven.
- Lead with what I do now, not where I studied.
- Version C may use fact 8 once. A and B must not.

The two instructions doing the most work are use only what is in the facts and no adjectives about me. The first blocks invention. The second blocks the register that makes bios unreadable, because almost every bad sentence in a bio is an adjective about the subject dressed up as information.

Wrapping the facts in tags rather than pasting them loose matters more than it looks. It gives the model an unambiguous boundary for the only-use-this rule, which is the same reason we recommend it in using XML tags to structure prompts.

Step 3: the attribution pass that catches inflation

This is the step people skip, and it is the one that makes the output safe to publish. Do not read the draft and judge whether it feels accurate. Make the model show its work.

For version B, output a table. One row per sentence.
Columns: the sentence, the fact numbers it draws on, and any
part of the sentence not supported by a numbered fact.

The third column is the point. You will typically find two or three rows where it is not empty, and they follow a pattern: a six-year stint becomes extensive experience, leading a team of four becomes scaling engineering organisations, one conference talk becomes a regular speaker. None of those is a lie the model knew it was telling. All of them are claims you would have to defend.

Fix them by editing the fact sheet or deleting the sentence, then regenerate. Do not negotiate with the draft.

The same forced-attribution trick is the general answer to a model asserting more than it was given, and we go further into it in how to stop AI from making up citations and prompting AI to flag what it is unsure about.

Step 4: make it sound like a person

The bio is now accurate and probably still slightly stiff. Two options, in order of effort.

The cheap one: give the model two bios you admire, from people in your field, and ask it to match their sentence rhythm and level of formality without copying any phrasing. Examples move output more reliably than adjectives about tone, which is the whole argument for multishot prompting in the vendor's own documentation, and we covered the question of how many examples a prompt needs separately.

The better one: write the last sentence yourself. The opening is formulaic by convention and the model handles it fine. The closing line is where a bio either sounds like a person or sounds like a template, and it is one sentence. Fact 8 in the example above exists for this. Restores mechanical typewriters is a real detail that makes a reader remember you, and no model would have invented it.

If you maintain a house style across your site, feed it in rather than re-describing it each time: see prompting AI to match your brand voice. The broader set of techniques behind all of this lives in our prompt engineering pillar.

Keep the fact sheet, throw away the bio

The output of this process that has lasting value is the numbered list, not the paragraph. Bios go stale in a specific and embarrassing way: a funding round closes, a cofounder leaves, the company pivots, and the bio on the conference site still describes the old thing eighteen months later.

So store the fact sheet somewhere you will find it, dated, and set two triggers for revisiting it. One is calendar-based, twice a year, which takes four minutes because you are editing numbered lines rather than rewriting prose. The other is event-based: any time something on the list stops being true, or something happens that would belong on it.

When you regenerate from an updated sheet, run the attribution table again. The second generation is where inflation creeps back in, because the model now has more material and more room to characterise it. A fact sheet that has grown from eight lines to fifteen will produce a bio that wants to summarise rather than state, and summarising is how led a team of four becomes built the engineering function.

Common failure modes when you prompt AI to write a founder bio

Frequently asked questions

How long should a founder bio be?

Keep three on file: around 25 words for listings and panel introductions, around 60 for your own site, and around 120 for applications and longer profiles. Generating all three from one fact sheet in one pass is what keeps them consistent, which matters when two of them end up on the same page.

Should a founder bio be first or third person?

Third person for anything someone else publishes, such as a speaker listing, a press mention or a company about page, because that is the register those contexts expect. First person for applications and personal sites. This is why the prompt above asks for both rather than making you pick.

Will AI invent credentials I do not have?

It will inflate the ones you do have, which is the more likely and more dangerous version. A six-year role becomes extensive experience without any intent to deceive. The attribution table is specifically there to surface that before publication.

Can I use the same bio everywhere?

The 60-word version travels well, but a funding application and a conference listing genuinely want different things, and the cost of the variants is nothing once the fact sheet exists.

What if my background is unimpressive on paper?

Then specificity is your only asset, and it is a real one. Shipped a tool that 40 bookkeepers use daily beats any adjective, and it is the sort of line a generic prompt will smooth away if you let it.