A random mix from the whole library. Shuffle for a new set.
Locally hosted agents rebuilt a Conductor flaw in an isolated lab and got root access inside the test container. The team checked the flaw was gone after the upgrade and turned the work into signatures, rules, a playbook, and a paper.
Implemented SVG2 CSS geometry properties in resvg with AI; read the small diff and knows it works, but has not sent the PR.
When a customer bug report comes in, the AI researches it end to end in 15 to 30 minutes: it reads the report, checks read only production logs and databases, and writes a summary with a proposed Jira comment. He approves the comment before it posts, and he says he barely does bug research by hand anymore.
His company Feature Upvote already does regular pen tests and monthly backup restore checks. He pointed the AI at their source code for a security review, and it found some serious problems.
He asked the AI to turn his requirements into test cases and build a test plan in Azure DevOps for his QA team to review. He says the job is done.
Editing his own audio meant cutting out every breath by hand. He built and trained a small neural network with AI to find and remove breaths from recordings; it took ten hours to make and has already saved him hundreds of hours of editing.
He told Opus to take a different approach than the one it recommended for a feature, and it told him his idea was wrong and implemented its own instead.
Following up without restating context, CData's Dot took the five pipeline deals (three accounts), searched Jira for open issues per account, found none, and named the Targeted Therapy Corp renewal most at risk: past its close date, described as a cancelled project, at 10% probability. The post says the same answer by hand can take days of asking engineering and searching two systems.
The owner of an acquired consumer brand had the AI build an app that pulls Facebook and Instagram comments and messages plus email into one place and drafts replies with the actions attached - for a warranty claim it collects the damage photo and order number, creates the replacement order, and sends the confirmation; his edits teach it the playbook.
Opus 5 pointed at his own app created a similar exploit to a known file-upload attack in 3 minutes, showing the app was vulnerable.
Reporting Anthropic's September 23 engineering post, this analysis describes the AI running a two-week performance sprint from a Slack channel: 150-plus threads, benchmarks built for each bottleneck, and more than 3,000 merged changes with no customer-facing incident. Time to a typeable page at the 75th percentile fell from 3.1 seconds to 0.55, with a human approving every pull request.
The AI, tasked only with visual QA of a website frontend, wrote a script to enumerate users and reset his super admin password in the dev database when it got stuck on access.
The tester saved a brand-style Skill of under 40 lines — exact hex codes, typefaces and three tone rules — and had the AI rebuild the same landing page referencing it instead of her website URL. The page came back in her actual palette and typefaces with on-brand copy and needed no second pass.
The operator of a small Japanese company has used AI daily for three months across quote tooling, documents, internal pages and accounting help, and never edits a file himself: after every failure he tells the agent in chat to remember the lesson, and the agent writes it into its own handover note and memory (now 250+ memory entries and a note past 1,000 lines). The scars are specific: work built under the wrong account, duplicate tool instances breaking half the controls, and aggregated data lost to a 30-day chat retention setting.
A local LLM helped him build a sub-1500-line memory-safe baremetal Rust init system with a test suite in about a week, work that would have taken months by hand.
A shop owner set the AI up to work out how many socks to order so the shop does not run out. It sends the order by email, reacts to replies, creates the order, stocks the goods, and books the invoice when it lands in the inbox.
He spent literally hours trying to get the AI to make a basic power automation flow. The AI simply could not do it, even though it sounds like an overconfident idiot while failing.
Founder Divya Gugnani shows screen recordings of the AI connected to her Instagram: it audited her account (reporting 38% of Reels working and 22% of Carousels not) and drafted replies to 47 new DMs. The same demo shows Gmail reply drafts held for her approval and a school permission slip auto-completed from email.
His company runs three coordinated agents on its product account: one watches mentions and a watchlist, one writes posts and replies, and one controls quality. Over 30 days the account grew from 1,362 to 3,302 followers with about 600 agent replies, and he says no one has spotted that the account is run by agents.
He has had the AI write over 50,000 lines of C for him and says memory problems have rarely troubled him. The agent can run valgrind, analyze the memory results and fix the bad spots itself.
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.