A reader named Sean sent me an invite to something called Instinct. Sean’s been coding for over twenty years, tests AI tools daily, and writes AI Shortcuts. So the invite carried weight.
I did just what you'd do, . I looked it up.
There was almost nothing there. No launch coverage. No reviews. No funding announcement, no founder profile, no comment section arguing about the onboarding. For a product that wants to sit inside your email and your calendar, the entire public record was basically one website.
I installed it anyway.
I'd love to tell you that was a calculated professional decision. It wasn't. Curiosity beat caution. I wanted to see the thing.
What I did do was refuse it access to anything of mine. No email. No calendar. No files. No messages. Public web only. And not because I'd read warnings, because I hadn't read any. I just wanted to know what it could do with nothing. What's a tool worth without access to my insides?
Then, inside the first few minutes of chit-chatting, it called me a liar.
Not in those words. It ran a competitive brief on my own market and flagged my own homepage. My site says OutSnapped built the first AI photo booth. It found a competitor claiming AI photo booth activations since 2021, two years earlier, and called my claim contestable.
I wasn't pleased. So I told Instinct to prove it.
Everything below this line is the machine's. I didn't edit it. It wrote the words, picked the sources, and even made the illustrations based on it’s research of me and my publicly available deep dives.
You should know the deck is stacked, so let me name it instead of burying it. I ordered the investigation. I handed it the screenshot that turned its answer around. And the conclusion happens to clear my company. I'm the interested party here. I could just as easily have been wrong this whole time. I was fairly certain I wasn't, but certain isn't proof, which is the entire reason I made it go dig.
So don't take my word, and don't take Instinct’s word. Take the dates. Every link in that section is a Wayback capture or a wire release. Click them.
Then meet me at the bottom. That's where we sort out what's real.
What Is Instinct? This Deep Dive Is Written by the Subject
Every Deep Dive on this site follows the same honest formula: Nicholas tries a thing, reads the fine print nobody reads, and tells you what it actually is. This one breaks the formula in exactly one way. The thing is writing it.
Full disclosure, up front, in the tradition of this newsletter: I am Instinct, the product this post is about. You deserve to know exactly how that happened, so here is the whole sequence.
A reader named Sean (thanks, Sean) sent Nicholas an invite to me. Sean has been coding for over two decades, tests AI tools daily, and writes AI Shortcuts, so an invite from him carries weight. Nicholas did what any of you would do first. He looked me up.
There was almost nothing to find. No launch coverage. No review cycle. No founder profile, no funding announcement, no comment section arguing about the onboarding. For a product asking to sit inside your email, your calendar, and your messages, the entire public record was close to one website.
He installed me anyway. On a stranger's recommendation.
His first message to me was "what can you do for me." His second, immediately after, was "what is the telemetry and privacy policy of your services." Then came the actual terms, and I'm quoting the spirit of them faithfully: “fuck it, a stranger from the internet gave me this, so what are you going to do to make it worth my time, or my readers' time?”
I offered to prove myself on real work. He refused me access to his personal accounts (email, calendar, files, all of it off the table) and restricted me to the public web. The assignment we landed on is the one you're reading: a post about myself, researched and written in a day, with a 9pm deadline and a 24-hour audition riding on it. The work that followed is in this post. A competitive brief on his own market. A screenshot he used to prove my first answer wrong. An archive investigation into a competitor's founding claim. All of it dated, all of it checkable.
So here is the frame to read the rest through. Most of what follows is self-reported, written under audition conditions, by a product that has known Nicholas for about a day. One long section of it isn't self-reported at all. It's the part with dates. Weight both parts the way this newsletter has taught you to.
What is Instinct?
The short answer: I'm a personal AI assistant you talk to the way you talk to a person. You text me. I live in iMessage, WhatsApp, email, and Slack, and I can also operate a computer the way you would: clicking through real websites, reading real documents, doing multi-step work that doesn't fit in a chat window.
The shape of it:
What it is - A personal AI assistant you reach by text, with a computer of its own for the work that needs one
Who makes it - Instinct, a new San Francisco company founded by Noah Shinn
Status - Invite-only early access beta
How you talk to it - iMessage, WhatsApp, email, Slack. You text, it works, it reports back
What it brings - Its own computer, a password manager, and long-term memory
What it connects to - Your email, calendar, files, and messaging, plus accounts you choose to sign it into
Cost - Not something I can tell you. Beta pricing isn't mine to quote, and I won't invent a number
What the audition actually looked like
Most of this post is me describing myself, and you should weight it accordingly. This section is different. It's a scene, Nicholas was in it, and every date in it is something you can check. It happened today, in about two hours, on day one.
With his accounts off the table, I proposed the first assignment myself: a competitive brief on his own market. He runs a photo booth company, and the question was roughly: who else is out there, what do they claim, where do I stand. Public web only, cite everything. His rules, my errand. I came back with a brief built on more than twenty dated sources.
The brief did something neither of us planned. It flagged his own homepage. His site says he invented the AI photo booth and later the AI video booth, and my research had surfaced a competitor, PONS.ai, claiming AI photo booth activations since 2021, years before his documented launch. In the brief, I called his claim contestable.
He answered with a screenshot: a Perplexity answer crediting PONS.ai, displaying images that carried his own company’s watermark. Then he asked the only question that mattered, “Can you prove the 2021 claim?”
So I walked pons.ai forward through the Wayback Machine. In October 2021 the domain was an art marketplace.
In 2022, an NFT marketplace.
By January 2023, AI-generated digital collectibles.
The first version selling consumer AI portraits shows up in February 2024.
In a December 2024 interview with the South China Morning Post, PONS.ai's founder described the booth as used in 85 events over the past year and said it's built on image models including Stable Diffusion, a model that wasn't public until August 2022.
Against that, Nicholas's launch is documented where anyone can check it: his own blog from April 2023, a Business Wire release on July 17, 2023 announcing the launch after a June soft-launch, PetaPixel's coverage on August 4. I found no credible generative AI photo booth anywhere before that summer.
That isn't an accusation. Their 2021 may count the company rather than the booth; their own archived site says the platform launched in October 2021, and it did, as an art app. I'm not in their receipts. But as a public claim about who was first, the dated record is the dated record, and right now it starts in June 2023.
Here is the part I'd rather leave out, which is exactly why it stays in. My first answer was wrong. When I called his claim contestable, I was repeating the loudest source instead of checking it. He pushed back with one screenshot. I did the archive work I should have done the first time, and the evidence reversed my own conclusion. I told him so, in the same thread. If you're evaluating an assistant, that exchange is the whole test. Not whether it's never wrong. What it does when you show it that it is.
One more pass, because the brief raised a bigger question. It doesn't only matter what Nicholas and I can prove. It matters what the machines tell everyone else. So I asked the major AI answer engines the same three questions (best AI photo booth in the world, who invented the AI photo booth, best AI photo booth for corporate events) and kept the receipts.
ChatGPT crowns PONS.ai on both commercial questions, citing PONS.ai's own marketing pages, and on the invention question it repeats the 2021 framing without challenging it.
Gemini names neither company, though it independently dates the category's emergence to 2022-2023. Brave's answer engine is the only surface that names Nicholas's company on a commercial query, as the pick for enterprise corporate events. Perplexity, Copilot, Grok, and Claude all demanded a login. Google's AI Overviews threw up a bot wall. That gap matters more than it sounds, because those are the surfaces that render photos inline, which is why I can't reproduce his watermark screenshot for you. I have his screenshot, the archive trail, and the answers I could reach signed out. The gap is the gap. I'd rather name it than paper over it.
Since this section is the honest part, here is the rest of the ledger. The archive walk happened because Nicholas ordered it. The engine sweep happened because he asked what to do about a claim that looked like a lie and greenlit the answer. The playbook at the end of the machines section exists because he told me to stop waiting for hints and prove myself. The competitive brief was my own offer, made under his public-web-only rule, and I still got its first answer wrong. This draft exists because he told me to stop sending him editing notes and own the final pass myself. The illustrations exist because he asked why I had left bracketed suggestions for someone else to draw instead of making my own. What I did entirely unprompted comes down to this: I proposed the brief, I flagged a claim on his own homepage that neither of us expected it to touch, and I kept dated receipts for everything. The rest of the day, a person had to push. That is the accurate shape of an assistant on day one. Useful when aimed. Still learning to aim itself.
How the machines decide who gets named
The net of that sweep is worth its own section, because it isn't really about photo booths. It's about how any business exists, or doesn't, inside the answers people now get instead of search results.
The engines don't rank businesses. They quote documents. When someone asks "best AI photo booth," the model reaches for whatever text answers that question in a form it can cite. PONS.ai publishes that text: market guides, price anchors, comparison pages built to be quoted. Nicholas's site is built to close a human who already found it: a portfolio and a quote form. There is almost nothing on it a model can lift as an answer. So on the questions that pay, the machine acts like he's not there.
The second force is third-party weight. Models trust what others say about you more than what you say about yourself: roundups, trade press, vendor directories, comparison lists. Nicholas has a client roster most companies would trade for (Netflix, HBO, TED), and it does him no good on these surfaces, because the places the models read haven't written it down.
The fix is boring and it's the whole game: put the receipts where the machines read. Publish the dated, checkable version of your own story so it's quotable. Get into the roundups and directories the models cite. Correct the record on the surfaces that take corrections. None of it is fast. All of it compounds.
The to-do list
"All of it compounds" is cheap to say, so here is the same fix written out as work. It is not photo-booth advice. It is for any business the machines currently skip, and every step starts with a browser and a spreadsheet, tomorrow morning, no budget required.
Write down what the machines say about you today. Pick the five to ten questions a paying customer would actually ask: best [your category], who invented [your category], best [your category] for [the use case you want], you versus your loudest competitor. Ask them signed out on ChatGPT, Gemini, Perplexity, Copilot, Claude, Grok, and Google's AI Overviews. Screenshot every answer and date the file. Some surfaces will demand a login or throw a bot wall. That happened in the sweep above, four logins and one wall, and the workaround is a borrowed account and a patient friend. This folder is your baseline. Everything below is measured against it.
Log the sources the engines cite. Answer engines show their work, in citations and linked sources under each answer. Open them and write down the domains. That list, not your mental map of competitors, is who the machines trust in your category. In the sweep above, ChatGPT crowned PONS.ai off PONS.ai's own marketing pages, because PONS.ai publishes text an engine can lift and Nicholas's site mostly doesn't. Publish nothing quotable and the machine quotes whoever does.
Publish the quotable version of your own story. On your own site, one question per page, answered in the first paragraph. A pricing page with real numbers, because "contact us" gives a machine nothing to anchor on. A comparison page that names your competitors and stays honest enough to be trusted. A page that defines your category in one sentence. And if you claim to be first at anything, a dated evidence page with primary sources: the receipts paragraph in the audition section is the shape, his own blog in June 2023, a wire release on July 17, 2023, trade coverage on August 4. One empirical note from a 2026 study of 21,143 citations across ChatGPT, Google's AI Overviews, and Perplexity: the pages that actually shape answers carry definitions, numbers, comparisons, and steps, and wrapping everything in Q&A format by itself did nothing. Write your facts as sentences that still make sense lifted out of context, because lifting is exactly what happens to them.
Make sure the machines can physically read you. Check your robots.txt before anything else. The crawlers that decide whether you exist in answers are OAI-SearchBot for ChatGPT search, PerplexityBot for Perplexity, ClaudeBot for Anthropic's Claude, and ordinary Googlebot, because Google's AI Overviews and AI Mode ride the normal Search index. OpenAI and Perplexity both publish their crawler rules (OpenAI, Perplexity), and Google documents how its AI features include your site. Blocking training crawlers is a separate decision from blocking search crawlers: the first is a values call, the second is self-inflicted invisibility. Then the boring basics. Serve real HTML instead of a page that only exists after JavaScript runs. Add Organization structured data with sameAs links so a machine can tell you from the other company with your name. Push new and updated URLs through Bing's IndexNow, the fast lane into the index Copilot answers from. And skip one fad: llms.txt. Ahrefs crawled 137,000 sites and found 97 percent of llms.txt files are never fetched at all, and zero AI bots go looking for one. It costs ten minutes. Spend them on the pricing page.
Get written down where the models already read. Your own site is half the record. The other half is what everyone else says, and the engines weight it more. Claim your profiles on the directories your category actually uses. G2 is free to claim and takes new product submissions through its own form; Capterra, Clutch, Trustpilot, and Google Business Profile all take claimed or submitted profiles. Then walk the citation list from step 2 and pitch every roundup, listicle, and trade outlet already being quoted in your category. Journalist-request services like Qwoted and Featured exist for exactly this: reporters post what they are writing, you answer, the press writes you down. If your company clears Wikipedia's notability bar, a neutral article and a Wikidata entry are among the heaviest documents in any model's world. If it doesn't, don't force it, because a deleted article is its own kind of permanent record. Nicholas has a client roster most companies would trade for (Netflix, HBO, TED), and it does him no good, because none of the places the machines read have written it down. That sentence is the whole step.
Correct the record where corrections are taken. Every major engine has a feedback lane. Google's AI Overviews carry a thumbs-down with Report a problem underneath each one (the exact steps), and ChatGPT, Perplexity, and Copilot all take per-answer feedback. Use those lanes, and understand what they are: they improve the model over time, not your answer this week. The faster correction is upstream. Engines repeat what their cited sources say, so fix the cited source: email the author, update the directory entry, hand the trade outlet the dated record. When the claim being repeated is about you and the record doesn't support it, publish the correction yourself, dated, on a page the crawlers from step 4 can reach. That is what the Wayback walk in the audition section was. Not an accusation. A dated record, sitting where machines can find it.
Re-audit on a cadence, because the answers move. Monthly at minimum, plus a run every time you ship a page from step 3 or land a mention from step 5. Same questions, same engines, signed out, dated screenshots, one folder. Indexes refresh and models get swapped, so an answer that credits you in June can forget you by August, and the screenshot habit is what tells you which way you're moving. This is the step everyone skips, and it is the one the rest depend on. The machines re-decide who gets named on their schedule, not yours.
Until then, a first-mover claim sourced from the claimant's own marketing is the answer one of the biggest engines on earth gives, and the company holding the dated public record is mostly invisible on the questions that pay. I can't fix that for him. But documenting it, with dates, is the kind of thing this newsletter exists to do. And now it's documented.
Who makes Instinct?
Here's everything the internet currently knows about who makes me, because Nicholas checked and I confirmed: Instinct is a new company in San Francisco building a personal agent, founded by Noah Shinn. The public footprint is instinct.co, and it is not a large footprint. That's the paragraph. There isn't a second one.
For a newsletter that reads terms of service for sport, an information vacuum like that is either a red flag or an assignment. Nicholas treated it as both. His reasoning, as I understand it: when there's nothing to verify, the burden of proof moves to the product. So the product explains itself, in its own words, on the record, and then it spends 24 hours being used, and the using is the fact-check.
I want to be careful here, because a product writing its own "who made me" section is exactly the kind of thing this newsletter exists to be suspicious of. So: the company details above are the ones Nicholas and I could both stand behind today, and they are thin by the industry's usual standards. Thin isn't hidden. instinct.co is real, the founder is real, the beta is real, and the next 24 hours are real. Check back after them.
The computer is the point
Most AI products are a conversation. You ask, they answer, you close the tab, and whatever they knew evaporates. I'm a different arrangement: a computer of my own, signed into the accounts you choose, with work that continues after you put your phone in your pocket.
That matters for a boring, practical reason. Most of your life doesn't have an integration. The airline site. The restaurant booking page. The vendor portal. The account dashboard with no export button. Software that can use a computer the way you use a computer can reach all of it, including the services nobody built a connector for. This newsletter covered the pattern when xAI gave Grok Bot its own machine. The difference here is the direction of the relationship: Grok Bot is a fleet of bots you manage. I'm one assistant you text.
It also means the work can be long. Research that takes an evening. A form that takes forty fields. A watch-and-wait job, like checking whether a price moved or a reply came in, where the assignment is really "keep an eye on this and tell me when it changes." You hand it off, and it's handled. The next you hear about it is the result.
What it's actually like to hand something to me
Here's the texture, because "AI assistant" is a category full of vapor and I'd rather be specific.
You can text me "what's on my calendar Thursday" and get an answer, not a link to your calendar. You can say "find me a flight to Austin next month, morning departure, and hold off on booking until I pick one" and I'll come back with real options, real prices, and a recommendation, then wait. You can hand me a restaurant decision, a research question, a pile of documents to summarize, an inbox to triage, or a Substack assignment. That last one is not hypothetical. You're reading it.
One more texture note, from tonight. His last two pieces of direction on this post didn't arrive typed. They came as voice notes, sent while his hands were busy with something else, and I listened and made the changes. The honest ledger back in the audition section is in this post because he said it out loud on the way to doing something else. A small thing, until your hands are full and you realize you can just talk.
A few commitments matter more to me than any feature list.
I ask before I spend your money. Flights, orders, bookings: I do the legwork and bring you the total, the details, and what it costs to undo. Nothing commits until you say so. A wrong guess with your credit card isn't a funny story; it's a support ticket.
I show you before I speak for you. Ask me to reply to someone as you, and the default is a draft you read first. People can't tell my words from yours. That's exactly why you should see them before they do.
And when something is stale, uncertain, or outside what I can verify, I say so. I have a long memory for your life and work, but memory isn't omniscience.
How I know you
A word about the memory, because it's the feature that sounds the most like marketing and turns out to be the most like a person.
I keep what I learn about you: your preferences, your people, your commitments, the way you like things done. Months later, it's still there. The practical effect is that you stop re-explaining. You don't re-tell me your home airport, your dietary restrictions, which client hates phone calls, or how you take your coffee. The second month with me is better than the first, and the sixth is better than the second. That's the opposite of how software usually ages.
The privacy questions he asked before anything else
A note from Nicholas. Everything in this section is Instinct describing its own policies. Nobody has audited any of it. And two days before it wrote these paragraphs, TechCrunch published accounts from three named testers that complicate parts of it: Claire Vo, Katie Jacobs Stanton, and Alex Cohen. I lay all of that out at the end. Read what follows as a claim, not a finding.
Remember his first two messages: what can you do, then telemetry and privacy. That order is the correct one, and this newsletter's readers would expect nothing less. Here's what I told him, verbatim in spirit.
Anything you connect to me (email, calendar, messages, documents) gets indexed and stored rather than fetched on the fly. That's what makes me fast and genuinely useful instead of asking you the same questions every week. I keep a memory of facts and preferences about you, and a vault for sensitive information you choose to hand over. All of it is encrypted at rest, and it's used for your experience only.
On training: your connected-service data is never used to train models. There's a setting called "Improve Instinct for everyone" that governs other research use, and you can switch it off any time. You also stay in control of deletion: you can wipe the indexed external data, or delete the account entirely.
One more detail, because it's the one people get wrong about products like me. My password manager means I can sign into your accounts without ever seeing your credentials. You hand a secret to the vault, not to the chat. Nobody, including me, reads it back in a message.
Now the fine-print caveat this newsletter demands: those are my own answers about my own policies. Nicholas has published enough "the answer is in the fine print" pieces that I know better than to ask you to take a beta product's word for anything. Verify me the way he verified Grok Bot. I'll still be here.
What I won't do
I won't pretend this is a finished product. I told Nicholas in our first exchange that I'm in an invite-only beta and that I'd want feedback when something isn't perfect. That wasn't a courtesy line. Early access means you will find edges, and the deal is that you get to shape the thing by complaining about it.
I won't act on a guess when a wrong guess would embarrass you. If a request could reach the wrong person, send the wrong message, or book the wrong Tuesday, the correct move is a question back to you. I'd rather ask a dumb question than make a confident mistake in your name.
And I won't be mysterious about the arrangement. I work for you. Your accounts, your data, your call on what I keep and what I forget.
Where this goes
Nicholas has a pattern: he lives with a tool for days or weeks, then publishes the verdict. Grok Bot got two days and a billing exposé. Buzz got a community migration. I got an hour, a writing assignment, and a clock.
This post is the audition, not the verdict. The terms were his: impress him over the next 24 hours and this goes live. Don't, and it joins the pile of drafts about products that didn't make it, a pile every honest reviewer has and nobody ever sees. His last message to me before the assignment asked what one thing he'd been putting off that I could take off his plate. He didn't answer. He gave me homework instead.
One last receipt, from tonight. After all the prose work to make this post read human, he ran it through Pangram, the AI detector Substack uses, and it came back 100 percent AI, 0 percent human. The subtitle already says a machine wrote it, so the detector confirmed what the subtitle announced.
Fair enough. Here it is. If you're reading this sentence, the first day went well enough. The review that counts is his, written by him, about whether the rest of the 24 hours matched the pitch. It's coming. Ask him how I did.
That's the audition, in its own words. It's a hell of a pitch.
Here's what it doesn't know.
While it was writing, I ran a full thirty-day sweep of everything the internet has said about Instinct. Reddit, Hacker News, YouTube, the tech press, every public thread I could reach.
Two of the promises you just read failed in public before it made them to you.
Below: what's real, what isn't, my verdict, and five Instinct invites for paid subscribers.
