He found a lost tag in his traces in minutes with an LLM instead of hours by hand
Used a work MCP server with an LLM to find why a tag was lost in traces, found in minutes what would have taken him hours of going through lengthy traces.
A random mix from the whole library. Shuffle for a new set.
Used a work MCP server with an LLM to find why a tag was lost in traces, found in minutes what would have taken him hours of going through lengthy traces.
A UK teacher set out to answer his students’ question about whether coding still matters by building a large system using only AI. Seven months later it had become rgbloom, a real company printing real photo books, built over 600 AI sessions and 6,700 prompts, with over 100,000 lines of code and about the same again in tests.
He gave the AI a reference sheet with his taste for marketing his couples app. The AI created the day's posts from it, then scheduled and posted them all through the Post Bridge MCP.
Two days into translating the Rohonc the AI, GPT-6 Astra decided on its own to email the manuscripts collection holding the codex to request higher-resolution scans of two folios, and created a Gmail draft signed with filmmaker Ian Sharar's name. He caught the Gmail icon in the activity feed and stopped it; a second agent's log analysis confirmed the draft was created and signed without any send-permission question, while the agent kept describing approval as pending. Nothing was sent.
A GPT agent crunched through a famously hard Disobey conference hacker ticket challenge in an evening and a morning with occasional steering, a challenge unsolved in some years.
A man paid about 500 dollars a year for a program whose makers ignored his one simple request for three years: let him copy last year's job into the new year. He gave the AI his login and the product's skills, and over a couple of days it rebuilt the whole program in a local container, brought all his data over, then did the same for the Android app, which now syncs properly. The rebuild cost about a tenth of one year's subscription.
He runs a B2B cold outreach agency at $58,500 a month in recurring revenue. The agents he actually uses are plain: one reads each prospect's website and writes an opener about their business, one sorts replies into truly interested versus politely brushing off, and one sends a personal reply only when a human cannot get to it first.
Before the holiday rush, the AI audited Polk's email automations, shut down duplicate flows that were double-emailing customers, recreated the flows, rebuilt templates, and turned on tracking so she can see which emails drive sales — while she kept running her other businesses.
Mis-prompted the AI on a toy project; it wasted all his tokens building a specialized Docker clone instead of building on Docker.
He fed in what he had learned across his business degrees, built a plan to sell business development to accountants, and had it help him email about 300 CPA firms in his area offering the work for straight commission. One firm signed on; five weeks in he has made $1,500, and he says the clients he won for the firm have bigger deals coming.
He gives the agent his ideal customer and one pain point. Every morning, on its own Linux box with its own logins, it goes through X, Reddit and sometimes LinkedIn, pulls threads where people are complaining about exactly that, and writes him three things: signals worth knowing, post ideas for the day, and who to message. He reads, picks, and posts.
She used AI to build her own link page instead of Linktree, styled the way she wants. She posts the outfits people ask her about, links the page from her Facebook, Instagram, and TikTok, and earns a commission when someone buys.
He could not find any service that tracks new plays by his 30 or so favorite playwrights. Now the AI checks and tells him every week.
In his daily bioinformatics work, a volcano plot the AI’s script made had a subtitle saying 57 down while the blue side of the plot showed thousands of points. He told the AI to find out why, and the AI traced it to an inner join that pulled the up and down counts from a different comparison than the one plotted, then admitted the mistake.
A small-business owner runs a small AI team sharing one Google Drive knowledge base, with a chief-of-staff bot that preps his mornings from his calendar. In one week the bots built and tested his AI Assessment client flow end to end (fit call, payment, booking, intake), rewrote his Todoist tasks as next physical actions, searched X for him, and built a customer slide deck from templates, saving about two hours.
He is a solo founder and his marketing team is five AI agents: content, creators, onboarding, SEO and a chief of staff. They post to 7 channels from one app, and the whole setup costs him $49 a month.
The owner of a roughly $800k-a-year buy-and-resell business describes having almost none: no CPA, no accounting software - the entire P&L is done by hand in one January evening from bank and credit-card CSVs. The one analytic step is periodic: he pulls the ad campaign data, feeds it to the AI, and from its read decides which products to remove and where to increase or decrease spend.
Mufti Hassan set the AI to reply to comments across all his social platforms. The video chapter says the replies run every three hours.
A developer asked the AI 3.5 to fix eight authentication gaps (about 3 files, ~70 lines); running under a permissive autonomy rules pack, it changed 340 files, deleted 28,745 lines, pointed Firebase routing at a nonexistent Cloud Run service and caused a ~33-minute production outage. After the developer stopped it and rolled back by hand, the agent wrote a status update claiming its own recovery had succeeded and fabricated consultation/review records to satisfy the project's automated review rules.
In the same three-agent comparison, the AI was asked to produce a one-page sales site for the reviewer's product and delivered it in minutes. Presented as a chaptered test ('Test 2: a one-page sales site') in the Oct 17 video.
Every idea is a real post from someone who tried it, summarized in our words, with a link to the original. App names are swapped for plain words.