Thomas Wight (@thomaswight) header
Thomas Wight (@thomaswight) profile picture

building with AI · tools, content & automation

Joined January 2016
5 free GitHub repos I'm saving for my AI agents. Testing one this week: 1. OpenCode: coding agent 2. Superpowers: one method for Claude Code, Codex, Cursor 3. anthropics/skills: starter skills 4. Langfuse: see every step your agent took 5. Learn Harness Engineering: free course
7
9
306
Thomas Wight retweeted
Replying to @corbin_braun
Fable 5.2 just for planning though ?
1
420
Yay
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges. They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
28
Thomas Wight retweeted
Replying to @rileybrown
When tf can I get this like dayum
1
19
Thomas Wight retweeted
Replying to @ducksays @bot
Until you use 10% of your usage to make one 🤦‍♂️
1
1
1
48
6 jobs my AI team does every weekday: 1. spots new AI tools getting real traction and checks if they demo well on camera 2. pulls the numbers on every post and flags anything beating my normal 3. finds creator posts beating their own average and works out why 4. turns those patterns into video ideas and rough scripts 5. turns the scripts I approve into shot lists and edit plans 6. reads what's taking off in AI overnight and brings me the best finds by morning nothing goes out until I say yes, and I post it myself. which job would you hand off first?
56
New era of business
23
Tell me I’m tripping but. grok bot max usage has been quietly expanded today. no ceiling on anything for anyone
28
I gave my bots one raw take from my phone. 33 seconds of me talking. they handed back a 31 second draft reel. 7 cuts, 42 captions. I didn't open an editor. it got most of it right. then one caption turned a shoutout into me calling myself a leader 😅 tomorrow morning I'm putting the whole breakdown up. what it nailed, and what I'd fix before anyone sees it. would you post a reel your AI cut without touching it yourself?
1
37
I'm building this in public: the bot team, the reel editor and the shared brain that keeps them in sync.
15
Claude is about to stop asking permission for every step. from next week Claude (@claudeai) runs in auto mode by default in the web, desktop and mobile apps on Pro, Max and Team. an automated check blocks risky actions instead, and you can switch back to manual anytime. the numbers behind it, from Anthropic's own Claude Code write up: 1. people approve 97% of permission prompts 2. in a test with 1,053 paid testers, humans caught 13.6% of dangerous commands 3. auto mode caught 89% number 2 got me. I still check every post myself before it goes out, and now I'm wondering how much of that is a real check and how much is habit. are you keeping auto mode on or switching back to manual?
1
1
76
worth knowing: Anthropic says that test data is for Claude Code specifically, and they still recommend checking high stakes changes yourself.
16
a study tracked 26,811 students in China for 30 months. once they started using AI, homework scores went up 18% and homework time dropped from 64 minutes to 45. then their closed book exam scores fell 20% within six months. the kicker: the ones who kept spending the same time on homework learned just as much. the research is written up on @voxeu. I run my content with a team of AI bots, so this made me check what I still do myself: 1. I pick the ideas. my research bot brings me what's moving in AI every night, I decide what's worth my take. 2. I make the call on every post. nothing goes out without my yes, and I post it myself. 3. I make the big changes. 6 posts in one day got about 50 views total. one list post the day before is now at 233. so I cut to 2 or 3 a day. 4. I check one board each morning before work, not a dozen chats. the bots do the legwork. the calls stay mine, because the day I stop making them is the day I lose the skill fr what's one part of your work you'd never hand to AI, even if it could do it?
1
21
the study is Strömberg, Lei and Wu, CEPR Discussion Paper 21577. 26,811 students in grades 7 to 12 in one county in central China, Sept 2022 to June 2025. about 4 in 5 of the AI users finished homework way faster than anyone else and took the biggest hit.
20
3 lines from the Claude Duet launch most posts skipped: 1. it checks your apps every few hours, not nonstop, because nonstop would eat your plan's limits. 2. it reads and drafts on its own, but asks before it sends, pays, signs or deletes. 3. Anthropic says it can still misjudge what's urgent. my bot team runs on a schedule for the same reason as 1. what's the one thing you'd never let it do without asking?
1
21
it's a waitlist research preview for US Pro and Max subscribers for now, so I can't try it from Australia yet. Follow for more.
9
on Wednesday I told one bot I want my reminders in Google Calendar now, not Apple Reminders. I said it once. to one bot. the next night a different bot on my team already knew. that's the shared brain. say it once, every bot hears it. what would you only want to say once?
1
23
I'm building this in public: the bot team, the reel editor and the shared brain that keeps them in sync. Follow for more.
1
11
this week my research bot read X's open source ranking code and wrote me a playbook. the surprise: in the scoring, a reply from someone you follow back is weighted 30x more than a like. it changed how I plan my posts. what would you change first?
1
1
36
I'm building this in public: the bot team, the reel editor and the shared brain that keeps them in sync. Follow for more.
1
10
Here's a job map audit for your AI Chief of Staff. 5 steps. It sorts the last 30 days of your asks into jobs and flags what you're still doing by hand. The full prompt is in the replies, one step each. Which job would come out on top for you?
1
15
Step 5: Suggest which one should become a saved skill first, and why.
1
1
That's the full audit. Follow for more AI team setups like this.
3