Middle-aged dad, late to the game, but so glad to be here. Banner art purchased with sats from @isolabellart
3378 sats received
Filtering by: claude · clear filter
Another wild Friday night with a new AI model. 😂 Anthropic Opus 5 came out today. We'll see how good it is.
The big unlock for me the past 10-14 days is using Codex for everything, but having it open up Claude Code in VS Code to do work I want it to do. Better than me going back and forth.
Browser control is still blowing my mind, especially for boring inbox tasks and admin tasks.
#ai #claude #codex #opus5
I probably have a very niche use case, but for my new business Perplexity Computer is a linchpin. It's simply is the best research tool. Claude Code and Codex aren't close. And for what it does it's not that expensive.
Tonight it's Codex as the main worker, Claude Code offering feedback to Codes, Perplexity Computer to research.
Very much looking forward to some Friday nights when I'm not working. Gotta make this new business go.
#ai #claude #codex #perplexity #claudecode
6/20/2026, 2 engagements
If you have even a small budget for #ai - personally or with your business - check out Perplexity Computer.
It's an orchestrator of roughly 20 different AI models. It does such a good job of running agents in parallel if you give it a problem.
For my new business, I have to do some really boring research by looking at public databases. There are other tools you can use, such as one called Firecrawl, that theoretically should be able to do it better and cheaper, but it simply can't.
Perplexity Computer will use the Perplexity Sonar models to do the searching, but then it will use the Anthropic's Opus 4.8 model to do the thinking and the orchestration. It seems to oscillate between those two models, and then it will always use the cheaper Anthropic Sonnet model for basic things.
If you had just $20 a month, it's best to have Claude or OpenAI.
But let's say you were going to spend an additional $10, $15, or $20 a month, you can buy Perplexity Computer creidts. Perplexity Computer is going to speed up your learning curve if you watch what it does in real time and sit back with a cup of coffee or a beer and just watch it work. It basically teaches you how these things use different tools in different ways.
Anyway, just a thought.
...and the frontier models are called frontier models for a reason...
#ai #perplexity #perplexitycomputer #claude #codex
6/5/2026, 1 engagement
I'm using an AI tool I made that's specific to my business, for a task I've had to do for years. It's so helpful and will take a two hour task down to an hour.
I think that in a year this is going to be the norm for a lot of people: custom tools that are the perfect fit for your business, but by you, with iterations along the way to make it better.
It took a lot of iterations to build it, and it costs some tokens, but it's great.
#claude #ai #codex
Some thoughts on AI as I work with it all day, every day, as an educator who isn't technial.
You have to both (a) play with these tools and (b) use them to make your current work either go faster or be done at a higher quality. Speed and quality for work, fiddling with them in between.
Claude is nice to you, so it's easy to think it's better than Chat GPT. It's a better writer, and a very good thinking, but Chat GPT is better at most things.
Because Claude is nice to you prompting it with "Don't tell me what you think I want to hear - be honest with me" is crucial. ChatGPT doesn't need as much of this. It's always honest, and give sober outlooks on things.
People think "AI can't write" and that was true in August 2025. In May 2026 that's silly. I write a weekly newsletter to 20k people. I dictate it via Wisprflow, Claud writes the first draft and then I tweak it. I have fewer unsubscribes than I did before I started doing this. It still takes an hour, but it used to take two.
It's best to have both the $20 Claude and $20 Chat GPT subscription if you can swing it.
Claude is a great copywriter. ChatGPT keeps getting better.
https://wisprflow.ai/ is a must.
Kimi models with
are good when you want privacy, but I can't run a business with those models.
I think and Claude are the two must haves if you do knowledge work.
Knowing when to use High and Extra High and Medium with Codex is important.
Claude Code starts to do dumb sh*t past about 65% of the context window. When I ask it at that point if we should write a handoff and start a new conversation it almost always says yes.
I slept on Claude Cowork and that was a mistake. It saved me a ton of time a few weeks ago when I needed quotes from a dozen book publishers. It wrote the follow up emails too, and then compared the quotes. There was zero reason for me to do that work. I found a better company than my current printers.
Browser control in Claude and Cowork is voodoo.
"Early in their career" lawyers and accountants are in trouble. In a month I'll replace my bookkeeper. This tax season I created a simple tool that took all my tax info and was within 1k of what my accountant said I'd owe.
Again, Claude writes very well. You can prompt it to never use em dashes.
I should have more skill.md docs, but the key ones I have a so powerful.
90% of the time I either start a Claude Code session with /plan or /superpowers.
I spent a ton of time with Openclaw in Feb and it was a waste, other than it got me comfortable using the terminal, and then I dove into Claude Code.
I worry that token cost for plebs will quickly be really high in the next year or two.
What are you learning? #asknostr
#ai #claude #claudecode #chatgpt #codex #plebai
For roughly a month I've been using Codex as an "auditor" for what I'm building in Claude Code. It's worked well, but it's a lot of back and forth.
I found the following plugin tonight from a YouTuber and it's perfect.
https://github.com/openai/codex-plugin-cc
Looking forward to using it this week to speed things up.
#ai #claude #codex
I spent just 60 min last night having #claude Cowork and their Chrome extension do a data entry task.
Note that everything I'm about to describe I could see on my screen and I wasn't doing anything...just watching.
Took a name from a google sheet
Went a website and looked up the name - and it knew what text box on the screen to put the name into (I had to have the tab for this site open)
If there were multiple entires for the name it found the right one
Took the three pieces of data from the name I wanted
Put the data in the sheet - but only after changing the data into the format I told it to
Moved on to the next name
It took 2-4 min per name. There were 17 names. This takes a human less time, but not half as much.
The big issue is that it times out after a few names. And essentially breaks.
You might be thinking, "You could do this with Claude code. You don't want to use Cowork for this."
I can't because you have to log into the site with the data - you can't scrape it. But Claude, after doing four of these said, "This is really time consuming - why don't we try to do this in Claude code."
Today it's too slow to have it do the 1200 hours of work I'll need done this year - I'll have to outsource that - but no doubt in a few months it'll do it faster.
I listened to a podcast with the developer at Claude who built Cowork - which they did in 10 days with Claude Code - and he said that while it's slow and clunky it's better than Claude Code was when it was released.
I hope it doesn't get too fast too quickly as I'd like to have a moat in the new business with a data set my competitors don't have. But 12 months from now they'll have it.
Not only does 2026 feel like a race, it clearly is a race.
#ai #claude #vibecoding #wildtimes
1/24/2026, 2 engagements