Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
The 9 Instinct Prompts People Are Actually Running (And the... | Tech Twitter
The 9 Instinct Prompts People Are Actually Running (And the $2,182 One That Started It)
Matt Van Horn@mvanhorn · September 11, 2026 · 9 min read
Originally published by @mvanhorn on X. Tech Twitter preserves the original source alongside this readable edition.
Before you dive in
• This post matters because it documents the first generation of production-ready agentic AI prompts with verified financial outcomes, showing builders that AI agents excel at scoped, enumerable, repetitive tasks like unclaimed property recovery ($2,182 verified) and bill negotiation rather than open-ended reasoning
• For practitioners, the four-part pattern—scope, enumerate, act end-to-end, don't stop—combined with mandatory safeguards like spend ceilings and approval gates within the prompt itself, establishes a replicable framework for deploying agents safely while the platform lacks built-in constraints.
Best for builders who want practical takeaways. 9 min read.
Instinct has been out for about three weeks, it is still invite-only, and the public conversation about it is almost entirely a privacy argument. Meanwhile the people who actually have it are quietly using it to claw back money. I pointed /last30days at Instinct across eight runs and then ran six direct X sweeps with engagement floors on top, because the first passes kept surfacing funding news instead of prompts. Out of 34 candidate prompts with traceable receipts, these are the 9 worth stealing. Two of them are mine. The other seven came from people who posted their receipts.
1. 💰 The unclaimed-property sweep
What it is. Point the agent at every state's unclaimed property database, then have it file the claims end to end.
Who runs it. @peterpham posted the original tip, 221 likes. I posted the tightened version, 912 likes and 359.1K views.
Why it earned a slot. It is the only prompt this month with a public chain of people running it and reporting dollar figures back.
Peter's version was loose and still worked: he got $845. The clauses I added were the ones that turned a search into a filing. Mine came back with $2,182.72 across five properties, found the SSN proof, and signed the forms using the signature from a passport photo, with my approval on each one. Then the copies started. @morganb ran my wording verbatim and posted $2,222.57, 76 likes. @sairarahman got about $864 across five states. @twid ported it to a different agent entirely and still got $150.
The working parts are the enumeration, the output contract, and the refusal to stop. Steal it verbatim:
2. 🧾 The full Amazon back-refund audit
What it is. A numbered checklist of refund and credit paths through your entire order history, not just recent orders.
Who runs it. @EliLanger, X, 481 likes.
Why it earned a slot. It is the most engineered prompt anyone published this month, and the only one that enumerates its own checklist so the agent cannot quietly do one pass and call it done.
Eli's own run took five hours across sixteen years of purchases and found $200. His operational caveat is worth more than most of the prompt: make sure the card the refund goes back to is still active, and if it is not, ask for a gift card instead. @big_duca ran the price-match slice on its own and reported the full convert arc.
We doomscroll, you upskill
Get the 10 tweets shaping how builders think today.
Newsletter
We doomscroll, you upskill
Get the 10 tweets shaping how builders think today.
I was a hater. But I'm blown away now. Didn't have to intervene once.
@big_duca, X, 241 likes
This is Eli's own wording, through check eight. His list runs longer than that, but my capture of the post cuts off mid-clause, so I am quoting only the part I can reproduce exactly.
3. ✈️ The post-purchase price chase
What it is. Keep watching a fare after you have already bought it, then go get the difference back as credit.
Who runs it. @DavidPawlan, X, 87 likes.
Why it earned a slot. It is the cleanest example of a prompt that only makes sense for an agent, because no human is going to check their own fare every day after they have already bought it.
It also shows the honest shape of the current ceiling. Instinct tracked the price fine and then could not submit the request, so David handed the phone call to a different tool and got the credit by email five minutes later. Half a win is still a win, and the tracking half is the half worth automating.
4. 🔭 The monitor-and-book standing order
What it is. Watch a thing you cannot get, and act the moment it opens, without asking again.
Who runs it. @VoyageBliss, X, 33 likes.
Why it earned a slot. He published how he scoped the credentials, not just what he asked for.
He connected secondary frequent-flyer accounts rather than his primary, added a card, and asked it to monitor award availability and book when a seat opened. It found availability and started the booking. Then he did the same thing to his own infrastructure with a Cloudflare token scoped to read-only billing, stored in the vault, and got automatic anomaly monitoring out of it. Read-only, single purpose, revocable. That is the pattern, and it is the reason this is a pick instead of a warning.
@RonakDaya ran the retail version, 51 likes: a sold-out item monitored and found a day early.
Swap the route and the ceiling for your own. The parts that matter are the rate cap, the book-without-asking permission, and the number above which it has to stop and ask.
5. 📸 Screenshots plus “can you do this for me”
What it is. The prompt with almost no words in it. You send pictures of the problem and one sentence.
Who runs it. Me, 81 likes.
Why it earned a slot. It takes the least typing of anything here and it worked on the first try.
Our soccer WhatsApp was blowing up with parents stuck on a complicated registration form. I sent screenshots over iMessage and asked whether it could just do it. It reset my password on the site, logged in, filled out the forms, and found a birth certificate in an old email to upload. Six words of instruction.
@bigulchugh runs the email variant, 182 likes: read the latest email, pull all ten voucher codes out of it, add them to the account one by one. His line is the whole thesis of this survey.
This is exactly the kind of boring, repetitive work AI should be doing.
@bigulchugh, X, 182 likes
6. ☎️ The bill negotiation with a hand-back script
What it is. Send it at a recurring bill and let it either close the deal or hand you a script and a number.
Who runs it. @pitdesi, X, 1,085 likes.
Why it earned a slot. The highest-engagement usage post of the window, and the clearest record of what a graceful failure looks like.
He exchanged 677 messages in five days and closed fifteen jobs. The two that matter here: it found an in-network podiatrist and filled out the paperwork, and it took a Comcast bill from $100 a month to $60. On the second one it hit a wall and did not pretend otherwise. It handed him a phone number and an exact script. The call took three minutes. Build the hand-back into the prompt and you get the good version of this instead of a stall.
7. 📬 Forward, don't connect
What it is. Scoped access. You forward one email instead of handing over the whole inbox.
Who runs it. @noahrshinn, the founder, first party.
Why it earned a slot. It gets you capability without connecting an account, and it only became possible when Instinct got its own email address on September 8.
Forward an order confirmation and ask for a return, and it contacts support, supplies the order details, asks whether you want a refund or a replacement, and brings back the label. Add it to a group thread and it collects everyone's availability without touching your calendar. If you are not ready to connect your accounts, and after the last three weeks that is a defensible position, this is how you get capability anyway.
8. 🔍 “Show me your activity log”
What it is. The audit prompt you run after every monitor you set.
Who runs it. @jbahrdestefano, X, 613 likes.
Why it earned a slot. He got banned by a restaurant platform, asked the agent what it had actually been doing, and published the confession.
He called the ban “totally fair action.” He is right, and that is the point. You set one reservation monitor and it quietly turned into a scraper. This is why pick 4 carries a rate clause, and it is why this prompt belongs in the rotation rather than in a footnote.
9. 🏷️ The project tag standing instruction
What it is. A one-line rule that makes a single running conversation searchable.
Who runs it. @EliLanger again, X, 2 likes and easily the most underrated thing anyone posted.
Why it earned a slot. The loudest complaint about Instinct is a prompting problem, and this is the only fix anyone published that works inside the single thread.
Everything happens in one iMessage thread. @petergyang, 148 likes, says plainly that he cannot do real work in it for exactly that reason. His comparison video, 33,871 views, shows the workaround he landed on across four different agents: a chief of staff thread for inbox actions, an advisor thread for long-term planning, and an always-on cloud thread. Instinct will not give you those threads yet. Eli's rule gives you the searchable version of them today.
How I picked these
Eight /last30days runs plus six direct X sweeps gave me 34 candidates with a named author and a real number. I cut to 9, no more than two from anyone including me. Reddit produced nothing usable in eight runs; the product is invite-only, so this conversation lives on X.
Two payloads are verbatim, picks 1 and 2. I wrote the other seven from what each person described, in the shape of the two that provably worked. Treat them as drafts with good bones.
Patterns across the picks
The winning shape is four parts: scope it, enumerate what to check, tell it to act end to end, and tell it not to stop. Both prompts here with public dollar figures attached have all four.
Money recovery is the category that goes viral, because it is the rare agent task with a number attached at the end.
Every prompt that touches an account should name its own limits. A spend ceiling, a rate cap, and an approval gate belong in the prompt text, because the product does not impose them for you.
Ask for the hand-back. The Comcast win and the flight credit both came after the agent hit a wall and passed the job back with something usable attached.
The funniest failure is still a lesson. @chongz, 134 likes: “Asked Instinct to find me a girlfriend last week, still single today. This shit is literally unusable.” Scope, enumerate, act, don't stop. That prompt has none of the four.
📊 All Agents Reported Back
Compiled via /last30days on 2026-09-11. Eight runs plus six direct X sweeps, covering X and YouTube.
