A random mix from the whole library. Shuffle for a new set.
He spent about 5 minutes picking the mods he wanted for GTA4. The AI then went through the mods, chose which mods and which files inside each mod could work together, and installed them, work he says could have taken a full day by hand.
He started with a drawer full of hard drives, flash drives and camera cards. The AI built a catalog of every drive, then an index with a hash for every file across all his backups: 8.1 million files, 3.6 million of them unique, going back about 20 years.
A maker had an LLM design a circuit board by running KiCad headlessly from a terminal, checking the design in a loop until it passed every rule check. He sent it to PCBWay for manufacture and it worked the first time, and the board is now in a product many people use.
He normally develops his astro photos in Siril. This time he handed the AI the raw stacked FITS file straight off his Seestar S50 and asked for three things only: flatten the background gradient, remove the noise, and stretch the image to bring out the spiral structure.
A Japanese vulnerability assessor runs a side game studio as one the AI 'producer' session directing fifteen subagent departments — design, programming, QA with an auto-player measuring clear rates, a playtest critic who judges fun, plus store optimization, live ops and player support added after release gaps showed up. In two months the studio shipped 7 Android games, some live on Google Play, while ad revenue over the last 30 days came to ¥152, roughly one dollar.
Used nit-picking history-teacher agents to evaluate a few hundred generated stories for accuracy, and pointed the AI at HN comments to aggregate issues and fix them with his OK; reported issues mostly fixed.
Joao Silva installed the Jev decision-model skill into his AI fleet without writing integration code: he pasted a plain-English instruction into a session and the agent copied the folder, found the right .env file, added the key and ran the checksum verification. The skill then passed his mandatory three-reviewer gate (security, QA, technical writing) before being committed.
When OpenAI released more than 700 mathematical preprints, a researcher wanted to search inside them, including proofs and bibliographies. With GPT-6 Astra he built a searchable catalogue that organises the papers by research area with clickable keywords, full text search, links to the originals, and each paper's Lean formalization status, and he put it on his website for public use.
A man set up his agent setup so the AI could ping him on Telegram when it needed hands. Working on hardware together, he got a message every 5 to 15 minutes telling him to move a resistor or patch cable in a set way, and following the instructions they got the hardware working the way he wanted.
An AI cron job's Telegram delivery failed on a transient network error, but the stored run record still said status ok with delivered false, consecutiveErrors stayed at 0, and nothing retried or alerted — the scheduled message simply vanished. The failure was only visible by digging into the run diagnostics; anyone watching the chat or the job's status saw a healthy run.
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.