I build real apps with AI and show the process. Kords: Mac dictation that never keeps your words. Founder, Launchcore. Building in public from Liverpool.
@jacewbuilds Naming the requirement is the bar, the file alone is a hint. I'll score my next seeded test three ways: right file, named requirement, exact gap in the diff. Only the last counts as a catch.
Want to do the same? Copy this into ChatGPT, Claude, Grok or whatever you use:
"I'm in the UK and I'm getting spam calls and emails. Help me find which sales contact databases and data brokers are likely to hold my personal data (e.g. Lusha, Apollo, ZoomInfo, Cognism, RocketReach, ContactOut, Kaspr). For each one, give me the official opt-out or removal page, and where there isn't one, draft a short UK GDPR Article 17 erasure request email I can send, using my details: [name], [email], [phone]. Then give me a simple checklist to track who has confirmed. Also remind me to register my number with the Telephone Preference Service."
Try it, then tell me how many companies had your data 👇
#AI #DataPrivacy #GDPR
14 companies were selling my mobile number. I'd never heard of a single one of them.
They weren't apps I'd signed up to. They were sales contact databases, quietly selling my name, email and number to anyone who wanted to cold call me.
Suddenly the spam calls made sense.
So I used AI to fight back. In one afternoon it helped me:
• work out which data brokers were likely holding my details
• find each company's opt-out or removal route
• draft every removal request and email
• track who had confirmed
14 companies. 14 removal requests. Most have already confirmed my data is gone.
In the UK you have the legal right to ask any company to delete your personal data, and they have one month to respond. Most people just don't know these companies exist.
This is the side of AI nobody talks about. Not replacing people, but giving ordinary people the time and tools to take back control.
@jacewbuilds A seeded miss is the right test: it turns a vibe into a clear pass or fail. Would you count a partial catch, where it flags the right file but not the exact ask?
@jacewbuilds Fair, and that honesty is rarer than a bug catch. My test: pick one old change, replay it against the asks list, see if the diff flags it. Worth a try on Xiaoqi?
@jacewbuilds Review on the asks is the underrated part. A changed requirement shows up in the diff before it shows up as a bug. Has that caught anything on Xiaoqi yet?
@easidotrun Keeping the control icon in frame at both swaps is the fix. I'd open on the MCP dashboard shot so the first second shows the payoff, then cut back to the login. Are you posting it as a short?
Needed a teleprompter for my shorts, so I built my own in a day.
Script scrolls, camera records, pick 9:16 or 16:9 right next to Record, save the take to Google Drive.
An AI coding agent did the build. I did the nitpicking.
What tiny tool would you build for yourself?
@jacewbuilds Linking every row is what turns a red test into a two-minute fix. I'd add the date each row was signed off too, so a later failure shows whether the ask or the code moved. Do you keep the list in the repo or beside it?
New Claude Code power, new way to break things faster.
How I'll try Code Mode before trusting it on my 20 AI bots:
1. Throwaway branch first
2. A plan file it checks against
3. Read the script before long runs
Opt in: CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running.
Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 https://t.co/Aj5ofLiK1y
@jacewbuilds Repo for the list, PR for the evidence is a clean split.
Do you link each row in the PR table back to its line in the list, or just the open asks?
@jacewbuilds List first means gaps show up as empty rows, not surprises.
I keep mine in the plan file so the agent ticks off each ask. Do you keep yours in the repo or the PR?
@easidotrun Showing who really has control is the bit most demos fudge.
A small badge that never leaves the corner would do it. Are you keeping the cursor visible during the takeover too?
@jacewbuilds A blank line as a reading map flips the whole review. Do you get the agent to fill that table in itself, or write the list of asks first yourself?
@easidotrun 'Your turn' is the caption that matters most, so I'd make it the biggest on screen. Hold it for a beat, or cut straight to the agent resuming?
@jacewbuilds First pass for the files, human pass for the behaviour. That split is right.
I'd add one thing: make it list what the ask implied but it didn't touch. Missing changes hide better than extra ones.
@easidotrun The handback is the shot that proves it.
If the agent's next click shows it noticed your manual step, viewers get session continuity in a second. Captions on, or letting the screen speak?
One AI bot ate 31% of my weekly usage on a single project.
The fix wasn't a better model:
1. Fewer, bigger tasks
2. Routines twice a day, not hourly
3. Bots only ping me for deadlines, money or approvals
Send this to anyone whose agents run on a usage cap.
@weswinder Reading the files first is the whole game. A skill is just a prompt running with your agent's permissions. Quick check before installing: search it for curl, fetch and .env. Takes ten seconds.
@ryan_doser13 No MCP, no place in the stack is a great filter. I just picked Riverside as my one recording app for shorts and remote guests. What are you having Claude do with it first: clips or show notes?