Good news, founder: the bar for "brutal AI feedback" is currently set at be mean, score it 42/100, post a share button. That's not a competitor — that's a layup. And it means the room is wide open for something that actually helps you build a better business.
That something is what we're making at THE ROAST: a cast of characters with depth, a heat dial you control, and a brand built to be on your side at full volume. Sharp, not sour. Cutting, but always toward the thing you'll be proud of next quarter.
🐉 "Be a council." → THE ROAST
Anyone can prompt a model to insult your pitch deck. The thing that actually moves your business is being argued with, by people who think differently. That's a cast, not a temperature setting.
The wrappers are doing their best (and that's the problem)
Most "roast my startup" tools are built in a weekend, and you can feel it. They're charming! They're useful for forty-five seconds! And then you close the tab and your business plan looks exactly the same. Three reasons — none of them about the model, all of them about product design.
1. One voice in three costumes
Tell a single LLM to "play three harsh critics" and you get one person doing three accents. Same cadence, same jokes, same objections in different fonts. It's improv night, not a board meeting — and your plan deserves a board meeting.
2. Famous-investor cosplay, no receipts
"A top VC would say…" Would they though? Without a corpus, a citation, or a signature on file, that's just vibes-as-a-service. You can feel the hollow even when the line is funny. Your gut knows. Trust it.
3. The score is a sugar rush
"42/100." Cool number. What do you do with it on Monday? A score is the energy drink of strategic feedback — peaks fast, crashes faster, leaves you thirsty. The work happens in the second question, then the third. A wrapper that ends after one shot can't ask either.
🥗 Character roast: chewier, but you grow muscle
What we mean by "characters with depth"
A character — the way novelists and screenwriters use the word — is a stack, not a sticker. Pile these layers and you get a voice that actually argues for something:
- Bones — worldview, wound, want, need. The Tired VC is tired because they've watched the same pitch fail fifteen times. That backstory is why their feedback lands.
- Influences — the books, films, mentors and rivals that shaped how this voice argues. Tracked explicitly, weighted, visible. We show our work.
- Corpus — for verified and historical voices, an actual library of things they've said or written, retrieved on demand. The voice cites a real position rather than inventing a plausible-sounding one.
- Refusals — what this voice won't say. The Economist won't moralise. The Lawyer doesn't do crowd-work. Refusals are the negative space that prove the character is real.
- Heat — a 1–6 dial that takes the same character from gentle to nuclear without changing who they are. Mild Economist is still the Economist. Spicy Economist is still the Economist, just with the gloves off.
Stack those and a strange thing happens: voices start disagreeing with each other. That disagreement is the product. It's the moment you stop hearing your own thoughts in a funhouse mirror and start hearing something useful.
Three tiers of voice, labelled honestly
Not every voice can — or should — be a real person. So we run three tiers, badged on every share card:
- Council voices — author-owned archetypes (The Sarcastic Friend, The Tired VC, The Operator). No impersonation, no estate, no lawsuit. Pure character work.
- Verified voices — real living people who have signed off on a snapshot of their voice pack. Their corpus is theirs. Their refusals are theirs. Their name is on the door.
- Historical voices — figures whose work is in the public domain (Sun Tzu first, more on the way). The system prompt enforces chapter citations and refuses modern jargon. No "let's jump on a synergy call" from the 6th century BC.
The wrappers don't tell you which is which because they don't know which is which. We tell you on the card, every time — because labelled feedback is usable feedback.
How to red-team your business plan with the cast
Here's the workflow we recommend on your first session. Forty minutes. Replaces an entire afternoon of solo doom-spiralling.
- Write your plan in 150 words: who you serve, what you sell, how you make money, why now. Resist the fifth bullet. The fifth bullet is where dreams go to die.
- Pick three voices that disagree on purpose — say, an Economist (numbers), an Operator (Tuesday-morning reality), and a Sarcastic Friend (pattern-matching to ten dead startups you've mercifully forgotten).
- Start at heat 3 (Sharp). Save heat 5 (Brutal) for when you're ready to hear the thing you've been avoiding. You'll know.
- Read the disagreements, not the agreements. Where the voices nod is where you're comfortable. Where they fight is where your plan is under-specified — and that's gold.
- Pick one objection per voice. Write a one-paragraph response. If you can't, congratulations: you've found next week's work.
- Export the best exchange as a roast card. Put it where you'll see it before your next investor call. Past You is now Future You's coach.
Why we built a brand, not a dashboard
Here's the unglamorous truth nobody admits: in this category, the moat isn't the model. The model is rented. The moat is the brand — and a brand is what wrappers like ideaimprove.ai physically cannot ship in a sprint.
- The Roast itself — a name you can say out loud without cringing, a charter you can read in a minute, a heat dial that's ours alone. Things that travel in screenshots.
- Humour as the carrier — share-cards, Roast of the Week, a tone that gets forwarded. Wrappers don't get forwarded. Characters do.
- Roast Goods — small, deliberate physical objects that turn a tab into a desk fixture. A notebook on your desk works for you 30 days a month. No retargeting required.
- The sign-off ledger — every verified voice links to a public record of who approved what, and when. Copying that requires building actual relationships with actual people. That takes years. Years are the moat.
These look like flourishes. They're the defence plan. Distribution + character + provenance is a stack the wrappers cannot ship next quarter, no matter which model they swap in.
🏰 A brand is a building you can't clone in a weekend.
What a serious roast actually looks like
Choosing a tool? Here's the honest checklist. A founder-feedback session that earns a place in your week should:
- Disagree with itself. Real councils argue. If your three voices high-five, you got one voice in three hats.
- Cite sources when the voice is verified or historical. "As you wrote in 2019" is the difference between feedback and fan-fiction.
- Refuse the off-charter request. A voice that will say anything is not a voice. It's a microphone.
- Move up and down the heat dial without losing itself. Brutal is a setting, not a personality.
- End with something portable — a card, a verdict, one paragraph you can act on Monday. Not a number.
- Be hard on the idea, never on the human. Non-negotiable. It's the line between feedback that builds and cruelty that just feels like content.
The short, sharp version
"Brutal AI feedback" is going to be a crowded market for about two years. Then it thins out fast. The survivors won't be the cheapest wrapper or the meanest prompt — they'll be the products that built characters founders trust, on top of brands founders recognise, with receipts to prove the voices are real.
That's the build. Bring your business plan, pick three voices that scare you a little, leave the score in 2024, and turn the dial up. We're rooting for you. Loudly. With notes.