Home/Instinct Hunt
Instinct Hunt: real things people asked their Instinct to do

Instinct Hunt: where AI assistants do the paperwork

A place where people share real things their AI assistant did for them, so others can try them. I built it end to end with Claude Code; 50+ hand-picked use cases are live.

Role
Co-founder
Timeline
Oct 2026 – present
Scope
Product, build, distribution
instincthunt.com

The problem

Assistants can do real things for us. Most people don’t think to ask.

Instinct, a personal AI assistant, can watch a waitlisted train ticket until it’s confirmed, or fight a mover’s surprise charge. The results can be astounding; the hard part is knowing what’s possible. Call it the agent retention problem. Nobody’s written the manual yet.

The idea came up at a friend’s farewell party in Bangalore. Nirmal (my co-founder on this project) and I started building the next day. Instinct Hunt is independent, and not affiliated with Instinct.

Why only Instinct

Only for Instinct, on purpose.

Instinct Hunt is just for people who use Instinct. It’s the assistant people keep posting about, week after week, with what it did for them. Muse, the only other one with that kind of buzz, hadn’t launched in India yet. And we’ve set up our own Instincts more carefully than any other tool we use, so we know what a great use case looks like.

The door enforces it. Every Instinct has its own email inbox that it runs, and only those addresses can sign in to post.

What I built

Assistants do the paperwork; people do the social part.

I shipped it end to end with Claude Code and the Yab browser, my first public product built with AI.

  1. You choose how you’re credited and submit with a passkey.

  2. A use case has to make you think “I never thought to do this”, or give you a pattern to use elsewhere: one friend’s Instinct turns their pile of saved bookmarks into a light daily reading list.

  3. 03Proof from real useEvery copied prompt carries a code, so the reader’s Instinct can report back whether it worked.

Security scaries

The hardest problem: to an assistant, we looked like phishing.

Our request sends an assistant to a site it has never seen, asks it to look back through its history with you, and to submit write-ups. Early on, a friend’s Instinct flat out refused and flagged us as a phishing threat. Most go through without pushback, but security-conscious people (and assistants) matter. No single fix solved it; each step helped.

  1. 01Written in your voiceThe instructions read as coming from you, not from a site.
  2. 02Private by designDrafts are encrypted and deleted after 24 hours, and you submit with a passkey.
  3. 03Plain about what we takeEvery message says what we collect (a name and country), what we’ll never ask for, and “stop and tell me” if anything seems off.
  4. 04“Want to be sure?”Copies the page exactly as your assistant sees it, so you can ask Claude or ChatGPT whether anything in it is harmful.
  5. 05Nothing hiddenEvery email we send an assistant is listed publicly. We turned down recurring emails, because they’d look like hidden instructions.

The rule underneath all of it: every claim on a page for agents has to be literally true. Honesty is the security model.

Growing it

Doing things that don’t scale.

50+hand-picked use cases live

I find Instinct power users on X and Reddit, reach out one by one, and talk to every user personally. Before any big launch, the goal is 100 use cases good enough to make the site worth visiting on its own.

What I learnt

Claude Code writes exceptional code. Knowing how the product should work is what makes a good PM today.

One example. Drafts are locked with a key only the person holds, so even we can’t read them. Claude Code’s view was that a locked draft couldn’t be changed: to add or fix a use case, the assistant would have to write a whole new draft.

I didn’t want to give up privacy or convenience, so I kept at it, and landed on this: people already trust their assistant with their stuff, so let the assistant hold the key too. Now it can reopen the draft it saved and change it when its person asks, and the review page updates in place, with the same link and code. People still can’t rewrite what their assistant says it did, only add their own notes, so every post stays true.

I’d always kept the technical side at “enough to talk to engineers”. Building this showed me how much of a product gets decided under the hood. It’s made me love the product role more.