Also built
Live, and not a BetLivePopcorn
Researched facts that pop up over YouTube while you watch.
Pop-Up Video for anything on YouTube, except the bubbles are researched. Click the popcorn on a video and it reads the transcript, researches the video on the web, and pins each fact to the second its subject comes up: the backstory, who the people are, what happened after, and where the video is wrong. Every card links its source. It runs on the AI you already pay for (Claude Code, Codex or an OpenRouter key), with no Popcorn account or server. What broke: before release, an outside AI review found that a malicious web page could have reached the local server and spent your plan. The server now refuses anything that isn't you. Setup is one message pasted to your coding agent.
- Runs on
- Claude Code · Codex · OpenRouter
- First facts
- ~1 min
- Python dependencies
- 0 (standard library only)
- Setup
- one message to your agent
LiveSwordFinder
Finds the ugliest strikeout swings in baseball, every day.
A live MLB product that ranks swords, the two-strike swings where a fooled hitter's bat dies out front. It reads every pitch from Statcast, scores each strikeout swing, and gives each one a share page, a place on the leaderboards and a clip wherever MLB still serves one. What broke: the first version ran across five services, and when the API disappeared and the database paused, the site went dark while a failing health check went unread. I rebuilt it with no database and no server. A daily job writes plain data files, so a quiet off-season can't take it down.
- Pitches read (to Sep 27)
- 716,792
- Swords ranked (to Sep 27)
- 2,399
- Servers
- 0
LiveWe Read Them
True-crime reels built from the record, not the retelling. Published as a Claude Code skill.
We read the FBI file, the depositions and the bodycam so you don't have to, and show what they say against what everyone thinks they know. First case: Gabby Petito, a 3:44 film and eleven vertical reels made in one day. The whole process is published as a skill anyone can install: three research agents, a facts gate where every quote is checked against the page image, FBI Vault files through the Wayback Machine, news clips cut on Whisper word timings, one-take ElevenLabs narration with a second voice that reads the record, and picture rendered in code. It's updated in the open as it improves.
- Version
- 1.3.1
- Reels, first case
- 11
- Full film
- 3:44
- Install
- /plugin marketplace add joewilsonai/skills
LiveOpen Feedback Factory
Feedback from inside your app becomes guarded pull requests, and each user is told what happened.
Users ask for things in a small widget inside the app. Each ask is scrubbed of emails, keys and tokens, grouped with similar requests, and either queued, held until enough people ask, sent back with a question, or blocked; requests that touch auth, billing or privacy go to a person. A worker hands the sanitized summary to a coding agent on its own branch, a pull request and CI follow, and nothing merges until a recorded reviewer approval and every policy gate pass. The user who asked sees the outcome in the same widget. The public repository is an early alpha; newer work isn't released yet.
- SDK on npm
- @open-feedback-factory/sdk 0.1.0
- Public code
- early alpha
LiveCanvaKilla
The banner maker that knows where every platform actually crops.
AI banners and profile pictures for X and LinkedIn. Every prompt carries the platform's crop rules, so the face, logo and tagline stay out of the avatar circle, the mobile buttons and the strips the platform trims off. A live desktop and mobile preview shows where they land, and the proof export draws the crop guides on the image. Built because Canva kept hiding banner text behind avatars, mobile follow buttons and the strips X clips off the top and bottom.
- X banner
- 1500×500
- LinkedIn banner
- 1584×396
- Profile picture
- 1024×1024
