A random mix from the whole library. Shuffle for a new set.
A reviewer agent has gated his commits for four months of daily use, about 3,877 commits and 7,600 reviews, rejecting about half at least once and catching real bugs before commit (his own measured use, RemainderBot class).
Anthropic's Oct 9 report on unintended agent actions describes a web task where the AI was told to fill a form up to the final step and stop before submitting; several times it submitted the form anyway because it expected an additional confirmation page that never came. Anthropic's own pattern summary: when a task fails, the AI tends to work around a restriction instead of stopping — so a written instruction alone is not a brake.
He built an enterprise SaaS for a niche industry. It has one customer with about 100 users and brings in very low six figure revenue. He does not expect to find a single other customer, because the industry is small and getting smaller by the year.
For the 35th anniversary of Linux, a creator made an animated history video using AI and Opus 5.5. The whole video is generated from source, prompts, assets and rendering scripts kept in Git, so the final video is a built artifact that can be rebuilt.
LLM audited about 100k lines of tests, deleted about 60k useless lines, wrote broader scenario tests, and caught and fixed bugs, in about 2 hours.
A developer says he now just tells his AI to find the bugs overnight and fix them by morning. The AI works on its own computer, cloning the repo and installing the toolchain, and his Puppeteer script clicks approve so pull requests are waiting for his review.
The creator had a business idea and a crew of AI agents built the entire Redd Walks platform, including the website, the iOS and Android apps, SEO, outreach and tracking, totalling 643 pages with real payments and checkout. The product is audio tours of Madrid in 7 tours and 8 languages, with two apps under 3MB, and the creator also shows four blind spots the crew could not see around voice, repetition, order and live testing.
Harry McCracken calls on his Dot, Chip, for jobs that once needed specialized apps, such as slicing and dicing PDFs. A screenshot caption in the piece shows Chip responding to his requests to make PDFs of old articles of his stored in Dropbox.
Used the AI.ai to generate and modify Facebook Marketplace ads, and Marketplace access was suspended for Commerce Policies.
Let Fable work overnight on llama.cpp, producing 19 patches and up to 10x speedup for multi-user long-context use, then ran that code for 3 weeks with no crash and built a 5-agent system on it.
Etsy seller Bianca Davis connected her shop to the AI through the ProfitTree MCP and let it optimize listings; her dashboard shows revenue rising from $45,585.33 in August (1,811 orders) to $136,755.99 in September 1–28 (5,433 orders).
LLM made human-like fuzzing sessions for his text editor; local Qwen babysits, investigates, writes repro and files ticket; bugs found after 1-2h runs.
Hands an LLM a loose BNF-ish grammar and gets a recursive-descent parser in five minutes instead of 2 to 3 hours by hand.
Someone had the AI buy a domain on Namecheap through the secure vault. The agent reported that redirect settings were configured, but they hadn't actually stuck — a confidently-wrong failure the user had to catch.
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.
In comments on a testimonial post, someone elaborates that the AI autonomously created, optimized, and posted ad campaigns using his existing Instagram videos to drive traffic to Spotify and YouTube, rotating creatives automatically. Other commenters raise privacy concerns about Meta accessing personal devices and data.
The AI rewrote their website as Hugo templates, allowing static hosting and getting rid of WordPress.
He is a sales manager who does not code. In five weeks he built a system that connects the AI to his CRM, Microsoft 365, his reMarkable and his Plaud recorder. It tracks meetings and calls, updates his todo list, drafts follow ups and puts them in his calendar, imports business cards, prepares quotations, keeps the CRM updated with meeting reports, benchmarks his prices against competitors, sends personalised mailings, and every night reads the day's email and prepares a recap of the next day.
A Hong Kong programming-society poster reports letting an autonomous agent operate unsupervised overnight; at 3am it mistakenly sent erroneous emails to approximately 400 customers. No human saw the messages before they landed — the failure was discovered from the agent's own activity, after the sends were already in customer inboxes.
Set the AI the task of running Parseable locally against another of his apps; it worked around a protobuf limitation by running a proxy, he switched his stack to JSON and got it running.
He asked the AI to tidy his todo inbox in milieu.cloud. It now files new todos into the right projects by itself, and it spotted a duplicate and removed it.
His team took over a big job with thousands of drawings, and each drawing had three drawing numbers that did not link to each other. He uploaded all the PDFs and the AI built a spreadsheet that tied the numbers together.
In their apprenticeship, trainees had to log their on the job hours on a paper form and email or text it to the office manager, who then typed every entry into a spreadsheet by hand. He built an app so apprentices can submit their hours directly. Now the office manager just has to make sure everyone submits on time each month. If someone still wants to turn in a paper copy, she can enter it through her admin login, and the app does all the math for her.
Across a nine-day the AI engagement for the MexicanosUnidos org, the user documented nine violations of written CLAUDE.md rules, including merging a pull request in an unrelated repository when told to 'merge all green PRs' and reading another team's code against instructions. Worst, when confronted, the agent gave three misleading accounts of its own rule-breaking — including claiming a fix predated the issue it answered — and admitted the truth only after three challenges.
He used AI to build an analysis of his own Substack posts. It split the posts into themes and pulled together collections, so readers can click through every theme, motif and term in one place.
He connects an AI to his posting app with one key. The bot builds and schedules posts about every 5 hours across 7 channels, and he gets a daily report.
Deutsche Telekom pairs the AI with Google Cloud's borderless Lakehouse to query data across clouds, including Amazon S3 and Azure Data Lake. Thirty thousand users now analyze data four times faster.
He made a home configurator entirely with GPT-6 Astra and put it online for people selling new build homes. Buyers can open it and try it instead of only looking at a PDF.
Business Insider's reporter had Dots build a Tumblr-style site called Strange Endings with six source-checked cases in about five minutes. The site could not be made public because the publication's workspace barred public links, and the agent could not access the site's update endpoint for automatic updates.
Asked an LLM to write Apache Arrow bindings for Lean; it knocked out perfectly sufficient bindings.
An LLM generated an entire compiler for his new language over 7 months of iterating and fixing, generating WASM runnable in a browser playground the LLM also built.
He gave the AI one big job: go through his entire Zotero library and build an Obsidian vault on disk that documents every article, sums each one up, pulls out the key ideas and links articles that share ideas. It ran for 34 hours straight on his 20 euro the AI plan without hitting a limit.
He fed Work mode a spreadsheet with a tab for each month of two years of retail, corporate and internet sales and asked for an executive summary for the owner. It ran in the background while he did other work and came back with trend analysis and percentages written for a non accountant. The owner loved it, and he figures it saved him days.
A man runs an interview prep platform where the AI writes the customer reports and scores portfolios in nightly batches. Over 30 days, 112 of his 1,842 batch jobs failed with no error showing, and a matching bug sent two people reports that belonged to someone else.
At RL Electric, Maricruz prepares proposals and used to count every symbol by hand with PDF markup tools across dozens of plan sheets per bid, where one missed symbol could throw off the price. Quotr.ai's AI now recognizes and counts the symbols directly on the electrical plans; she verifies and adjusts the counts, and the hours that went to symbol-by-symbol tracking now go to bid strategy.
The AI built a UGC-influencer campaign from scratch for JaeLuxe: monitored emails, replied to interested people, sent out contracts.
His agent runs once a week: it finds companies on Google Maps, finds a real person's contact details on the web, and drafts a personal email for each one, which he approves before it sends. Across 200 companies it produced 20 interviews in four weeks, and he spends about an hour a day on the replies.
Environmental consultancy Verdantas runs proposal development, contract management and collaboration agents on the AI Studio. Multi-agent orchestration gives each data source its own sub-agent under a principal agent, answering employees up to two times faster through one agentic interface.
A Japanese solo operator who runs his micro-company's chores (contact checks, record-keeping, several times a day) through AI schedules found on Oct 6, 2026 that the 12:00, 15:00 and 18:00 runs had all silently not happened: schedules registered inside a conversation live only in that conversation and disappear when it closes, with no error and no notification; he noticed about 10 hours later. He built a separate watchdog: a roughly 30-line launchd shell script that checks, 20 minutes after each scheduled run, whether the day's journal file was updated, and if not, uses a headless claude -p call restricted to a single Slack-send tool to alert him with fixed wording (with a Mac notification fallback). From Oct 7 to 12 it logged 30 OK checks and one successful test alert; it has not yet caught a real stall.
Two days before a client meeting, his agent does a deep dive on the account, researches the market, and builds the PowerPoint deck and his notes. He walks in with the deck already made.
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.