He works as a commercial roofing estimator for a large regional contractor in the NYC metro, and his work starts with large PDF construction drawings and specifications. Over eight months he built a system he calls Estimator-OS. Incoming drawing and spec sets get broken into the standalone sheets and sections relevant to his trade, filed and organized, then chunked, documented and tagged into a local database for the project. From there it runs his RFI process and addendum reconciliation and issues his pricing requests to vendors and subs. Before takeoff he gets an HTML dashboard of the project with all the facts, details and gotchas. He does the takeoff in the software they use, then gives the final number back to the AI for review and final checking. He runs it on a Max 20x plan and values the usage at about $6,000 at API prices, and says he moves through projects more quickly and with greater accuracy.
A five-agent the AI recovery desk searched the owner’s contracts, invoices, email and ledger for a $42,900 loss and found the signed deal, the approved invoice and the promise to pay. The collection draft stopped at the owner’s screen for his say-so, with the evidence chain already assembled.
$42,900one time
This only partly worked. Worth knowing before you try it.
In his new job organising a small business he built spreadsheets that work out full product costs from current material prices and product dimensions, including all costs of goods sold, so pricing a product or bidding a custom job takes about 8 seconds. Within minutes the sheets showed a raw material rise was about to eat a giant chunk of margin, and the business raised its prices for the coming year.
Creator asked her agent 'Everest' for a 5-day PCOS-friendly meal plan at 100g protein per day, Mediterranean style. It produced the plan, a grocery list, a Sunday prep schedule, and set a Saturday 6pm calendar nudge for weekly prep.
Creator someone ordered three items from clothing brand Plus44 while in Turkey in December 2024; the package could not clear customs for a tourist without a Turkish ID and was returned. After the promised refund never came and the 120-day chargeback window expired in April 2025, later emails went unanswered. He asked the AI for help: The AI identified the UK company (Plus 44 World Limited), cited Section 29 of the Consumer Rights Act 2015 and sent a strongly worded letter demanding the refund within 14 days or court action. Within two hours Plus44 issued the $478.39 refund. (Caption carries his AI referral code.)
Working with AI on his NeoBox emulator, he found setjmp was eating 81 percent of frame time, switched to _setjmp for a roughly 130x per-call gain, and the fix was merged upstream in Geolith.
J.t. Chen (posting in Chinese) asked the AI before work to win back a $20 home-internet discount lost when two phone lines left AT&T. The AI could not restore the same discount, but negotiated an alternative reduction of $5/mo with support chat. His main delight was avoiding the usual 30 minutes to 2 hours of waiting and typing with US customer service. In comments he clarified the AI works over web text chat, not phone calls.
$5
This only partly worked. Worth knowing before you try it.
A hobbyist says the AI assembled the firmware running his living-room lights and fans: an ESP32 web server driving a radio transmitter. It also built him a Streamlit app to explore the signals, and the code is public on GitHub.
A researcher had the AI systematically scan the mystery constants in the RC2 cipher and reverse-engineer how its 256-byte permutation table was generated. The derivation and reproducible code were published, with all 256 entries reproduced exactly.
On its first scan of the agency owner's connected Slack and calendar, the Dot Toby surfaced a Monday client discussion involving work outside the budget alongside a calendar overlap, plus another meeting conflict on Wednesday. It connected information across both apps without being fed anything manually.
Vicente Teng (42K followers) missed an early morning flight about an hour before waking up. Still groggy at 7:30 AM, he asked the AI to find a cheaper replacement.
This only partly worked. Worth knowing before you try it.
LLM wrote tests comparing his code to kernel syscall tables and wrote a tool tracing all kernel calls to generate a bare minimal hardened kernel config, saves him security review time.
Marshall Haas, a serial founder (140K followers; sold his last company for $52M), got early Meta access to the AI and set up an agent named 'Nova' with standing instructions to proactively handle his bills plus business and personal tasks. His caption: 'Meta gave me early access to someone AI...
His free French teaching site Ralentir travaux, used by millions of visitors, had been more or less abandoned because its old code ran to thousands of pages. Over about six months the AI and the AI cleaned, fixed and modernised the code and added new tools, and the rebuilt site relaunched for its 20th anniversary.
He gave the AI a CSV of old macro tracking data and asked it to add the data into the folder where he tracks it. The AI misread the spreadsheet, decided the new data contradicted what was there, and overwrote his existing data without asking.
This one went wrong. See what happened before you try it.
Meta's own example gadget puts the AI on a Seeed reTerminal color e-ink display. The AI shows a morning briefing, reminders and a shopping list on the screen, and the code to build one yourself is open source for anyone with the same display.
He built one skill that does his company credit card expenses and another that files his email in Outlook. Both were manual jobs he hated. Now he mostly watches, answers the odd question and finds an invoice when asked.
His local school division released the raw answers to a touchy redistricting survey as one PDF of more than 3,000 pages, which nobody could realistically read. He used AI to write scripts that pulled the answers out and built a public web page where people can search the responses and filter them by keyword and by school.
His 2019 LG TV has a great screen, but the official Plex app took about 30 seconds just to show the profile picker, because it is really a web page. There is no public native toolkit for the TV, so he rooted a spare set, read the TVs own software with analysis tools and the AI, and built a native Plex client of his own. On the same TV it shows the profile picker in about 3 seconds.
A reviewer asked the AI for a presentation with a single request and no detailed explanation; it produced the deck, and when she commented on how she wanted it to look, it revised the presentation to match exactly.
A poster who is neither a web designer nor an engineer redid the layout for a personal portfolio site in Figma, then built it with help from the AI and the AI. A working version took 30 minutes, the two agents reviewed each other's code while the poster tuned the look and the admin screens, and the site nagiko.me is now live.
Silva built a one-command routing gate on the AI that answers four typed questions about any request — which of his nine agent profiles should take it, whether it needs him first, how much work it is, and whether the text contains a live credential. On five real requests it flagged a key-shaped string at 0.94 probability and routed it correctly, at a median 437 milliseconds per gate call.
Àlex sent his Instinct assistant a screenshot of an Obeses t-shirt and told it his size. The agent found the shop, he confirmed the payment through a Shopify email sent via Link, and the t-shirt was at his home the next day.
He used AI at work the way he was expected to. He asked it to plan a task, answered its questions, and approved the plan. It ran for 2 hours, then said the plan was complete and that no source files were changed.
This one went wrong. See what happened before you try it.
Freelancer Justin Blaze Martin posted on X on Oct 19 that the first thing his Dot did was go through his three inboxes and find a $4,000 invoice from July that he had forgotten to send a client. The Dot assembled the invoice from the email thread and had it ready to send after his approval.
Noah gave his AI Q its own Google Voice number so it can text his partner directly. With his permission, if she asks something within reason the AI answers for him: it can tell her whether his phone battery is dead, or share his location, because he has given the AI location access. His standing policy is that anything the AI sends as him is drafted for his approval first.
An Italian user connected the AI over MCP to three sources of truth — Garmin for activity and recovery, Yazio for food, a FitDays scale for weight and body composition — and gave it a coach's rulebook in CLAUDE.md plus a git repository as memory, where every session ends with a commit recording decisions and rejected alternatives. Daily check-ins and a weekly review recalculate his real TDEE from actual intake and weight change, change at most one parameter at a time, sync approved workouts to his Garmin calendar, and stop everything for a doctor if red flags like chest pain or a ten-day HRV slide appear.
He asked the AI to help draft an email and expected a local markdown file. Because he had connected Gmail through MCP at some point, the AI searched his mail, found the right thread, and created the draft inside Gmail.
Built with agents over about a year: a production billing system for his small telecom company, a multiplatform app, internal tools in use, and a coding agent harness now used by non-technical colleagues.
A someone testimonial graphic quotes Barron's reviewer Josh Schafer: he uploaded his credit-card statement PDF to the AI, which flagged an Adobe subscription he did not realise he still paid for. Because the AI had no access to his accounts, it walked him through cancelling on Adobe's site himself, saving nearly $25 a month.
He asked his bot to sign him up for the side events around All-In Montreal. He woke up registered for 10 events, including some he did not know existed.
His road trip is fully planned by the AI, the destinations, the hotels and the whole schedule, with a backup option Fable 5.1 found in case the Badlands roads are rough. The route itself is driven by Tesla Full Self Driving while he follows the plan.
This only partly worked. Worth knowing before you try it.