logged on to X this morning to a message that my account has been locked due to suspicious behaviour. I verified my email account and it unlocked.
Then, it said that a label had been attached to my account due to suspected manipulation of X's services.
I dont get it. I type every message, I reply in good faith, the content is my own and not repurposed.
I built a browser tool in one evening, with Claude. It's called No Live. It stops every live stream on YouTube in the browser it's installed in. Live TV included.
Here's what made the difference:
Say why, not just what. I asked for a BBC blocker so people could say they don't need a licence. Claude checked the rules, said a BBC-only blocker wouldn't do that, and suggested blocking every live stream instead.
Test against the real thing. I ran the checks on real YouTube, in Edge, with a control run: the live stream plays without the tool, so a passing test means something. Playwright's WebKit build couldn't decode YouTube's video, so the main run is in real Edge.
Say what it doesn't do, first. It doesn't decide whether you need a TV licence. It doesn't cover the YouTube apps, other sites or iPlayer. The README says so near the top.
The full guide, with the first prompt I used: lifewithai.co.uk/tips/build-…
I hate editing video. So I built a tool that does it for me.
I record one take and drop it in a folder. It cuts every pause and every fluffed line (I just say "scratch that" and start again), adds captions, and hands me a phone version and a YouTube version.
First test: 39 seconds in, 18.6 out.
It even drafts the title and description, for about a tenth of a penny.
Going to start releasing more videos on X and other platforms.
If you pay for Claude Max, check your settings. There's free money in there most people haven't claimed.
Max now includes monthly credit for the Claude API: $100 on Max 5x, $200 on Max 20x. It doesn't cover Claude Code, and it doesn't roll over.
I moved my small background jobs onto it, the ones that were eating my Claude Code limits:
- my nightly digest of the day's work: about 3 cents a night (it was about 17 cents of allowance through Claude Code)
- a morning scout that reads my work and pitches post ideas: about 11 cents
- my video tool's titles and descriptions: a fraction of a cent
Two traps cost me an evening. The worst one: don't set your API key for your whole computer, or Claude Code grabs it and stops working.
Full setup guide in the reply.
How to claim it, what it covers, and both traps: lifewithai.co.uk/tips/claim-…
It's entirely free to use hillmade.uk/no-live
And for full transparency, I made it open source, so you can check the code and make sure I'm not spying on you (I'm not!) github.com/Hexakin/no-live
I set 93 AI agents loose on a game I'm making, with one job: find bugs. Don't fix anything, just look.
They came back with 35 confirmed bugs.
My favourite: hit restart after firing the Mega sheep, and the game handed you the Mega again. Free Megas, forever.
17 were fixed and merged within a day. The rest wait for the tuning pass.
Momentum feels good, but always makes sure to check your work at each stage, otherwise you are building on quicksand
The biggest issue I've run into learning to create with AI is knowing how to make it actually productive rather than just making me feel productive.
AI is an equaliser. It lets people create and ship things they couldn't before. But it's also very easy to chase the early buzz: start something new, feel momentum, hit a wall your own skills can't get around, watch the AI bounce off the same walls, and quietly abandon it for the next spark. I've done that plenty.
Over the last couple of years I've gone from a know-nothing office worker to someone who can make things people actually use. The biggest lesson wasn't a technique. It was stopping the hunt for that early energy and instead getting better at building my own knowledge, and at improving the tools I use so they compound rather than restart.
This weekend I gave Claude a research goal instead of grinding through tests and rewrites myself. Thoroughly investigate the next chunk of work, actively try to falsify the findings, and then build small proofs of the parts that hold up. The docs came back well-sourced, including citations that contradicted or complicated the obvious path. Claude was able to falsify where it could, and actually go ahead and build the parts that appeared well grounded, or those that literally needed the test to prove. I've reviewed several of the test builds myself and the work is great.
What actually stopped me was scrolling the log and seeing nearly 3,000 agent runs, using only about 25% of my 5-hour rate (I started the session already at 5% of weekly). Almost all of that capability is Claude Code and Anthropic. The last bit is just me having slowly learned how to set the work up so it stays useful instead of chasing the feeling of being useful.
I'm still not ready to show the project itself. The things I do share are on my site if you want to see whether any of this is real.
My good friend @jpshrodinger made a post the other day about a music video he got Opus 5.5 to make entirely with code (and it was epic)!
I went for an entirely different style, BUT, here is my own "code only" music video, for my latest single "Tinnitus"
let me know what you think, and lets share each others work, really wanna see what other people are creating







