AI, search, & video strategist | Monetizing my hyperfixations since 2012. Helping brands grow at https://nitter.cf/t.co/jL3wNHPY7w
Joined September 2022
- Tweets1.2K
- Following835
- Followers415
- Likes2K
Pinned Tweet
Skip Netflix tonight. Watch this 25-minute masterclass on how to build your first AI agent with Claude Code instead.
No code. No node trees. Just plain English.
What you'll learn:
- Why 95% of people using AI are leaving most of the value on the table
- The 4 ingredients that separate an agent from a chatbot (most tools only have 1)
- How to set everything up in under 10 minutes
- The "reverse prompt" technique that builds your agent's memory for you
- How to give your agent real tools — Notion, analytics, web scrapers — not just chat
- What autonomy actually looks like (I told mine to do keyword research. It got blocked, switched tools on its own, and got the data anyway.)
- 3 tips from months of daily use that most tutorials skip
I don't know how to code. Not one line. Built all of this by talking to it.
Full video below.
Me: Man, Opus 5.5 usage efficiency has really gone downhill.
Also me: *realizing I allowed a thread to spin up 12 Max agents all working concurrently*
Whoops.
Another self-promotional listicle spotted in the wild completely backfiring.
> Shiftflow writes a self-promotional listicle and positions themselves first
> Google AI Overviews cites the listicle as the first source, but ends up recommending a competitor instead
> Feels bad, man
@lilyraynyc
AI tip: if you find that your outputs are too overwhelming (like you have a question on something Claude said 500 words ago, but he's already moved on), say, "Give these points to me one at a time."
Been a total game-changer for me.
So all software is theoretically open-source now to those willing to throw enough tokens at it.
But at the same time, we live in a bubble. Only 4.5% of people even pay for a Claude/ChatGPT/Gemini subscription and most normies treat it like a better version of Google.
So how will this AI-assisted software cloning thing affect the software industry as a whole? Very slowly, and then all at once.
AI adoption is outpacing computers AND the internet, compute is getting cheaper, and AI is getting faster. Software companies ignore these trends at their peril.
I give my full thoughts in my latest YouTube video (link in first comments)
I feel like the "software is dead because we can clone it" conversation misses one of the most exciting/interesting shifts in how we engage with computers. Simply put, it's that we can now build only the specific features of the software we need and nothing else.
Imagine an invoicing tool that's fine-tuned to your business, your workflow, and your customer types. Or even dreaming a little bigger, a marketing platform or image editor or 3D modeler or even a freaking operating system that's custom built just for you. That's what makes gets me excited as both a user and seller of software.
I talk more about this in my newest YouTube video (link in first comment)
I'm pretty sure I've found the most optimal editing workflow for motion graphics/A-roll editing/B2B YouTube with Claude -> Resolve -> Remotion.
I've not been this excited about AI since the Opus 5.5 release (literally a month ago but it feels like forever).
Shout-out @vanschneider I'm quoting you in my video :)
Matt Kenyon retweeted
Desde hace muchos años desarrollo una especie de diógenes digital que se ha acentuado con la crisis de los 40. Este fin de semana he decidido poner un poco de orden al caos digital, y el resultado era una cosa impensable para mi hace un año.
Is the era of self-promotional listicles officially over?
A few days ago @lilyraynyc posted evidence of ChatGPT throwing shade on self-promotional listicles, basically disregarding them as biased and unfit as research on which to make a buying decision.
I encountered this in the wild when poking around ChatGPT for some client work:
1. Going straight to Reddit to gather firsthand experiences
2. Not even bothering to cite self-promotional listicles and saying "vendor sponsored comparison articles also deserve less weight than independent reviews."
Encountered in the wild. I queried ChatGPT for the best accounting firm management software, and it gave me this multiple-choice form to narrow down options:
As if @lilyraynyc and @glenngabe's articles on the matter weren't enough to convince you to kick low-effort AI content to the curb forever, consider this quote from Google's Quality Rater guidelines:
"The Lowest rating applies if all or almost all of the MC (main content) on the page is... auto or AI generated... with little to no effort, little to no originality, and little to no added value... Such pages should be rated Lowest..."
For those who don't know, Google's Quality Raters give sites 1 of 9 potential ratings, Lowest being the worst and Highest being the best.
According to this, if Quality Raters have even a whiff of a suspicion that your content is AI generated with no oversight, they're instructed to rate your site Lowest, end of story.
"Claude, write me an article on the best ways to mount a TV on drywall make no mistakes plz."
👆That's what Google's newest algo update is likely coming for.
Crutchfield Corporation does the opposite. High-effort content with original data, visuals, and "talent and skill" (one of Google's new positive content signals, if you can believe it).
This clip is from my new video on Google's "low-effort" content guidelines. Full video in the comments.
Matt Kenyon retweeted
Initially, I had Claude make this animation in 2D. It wrote the piano score and drew every frame in JavaScript.
I gave this 2D version as reference to another Claude. Because the whole film was code, Claude was able to use its exact choreography and make the 3D one in Blender.
This is what I’ve been saying. I have a standing rule in my CLAUDE.md that anytime I ask for browser or computer actions, Claude automatically sends that to Codex.
Hot take: 𝘁𝗵𝗲 𝗯𝗲𝘀𝘁 𝗯𝗿𝗼𝘄𝘀𝗲𝗿 𝗳𝗼𝗿 𝗢𝗽𝘂𝘀 𝟱.𝟱 𝗶𝘀 𝗖𝗼𝗱𝗲𝘅'𝘀 𝗰𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝘂𝘀𝗲.
I always thought it only worked inside Codex. Turns out it's just 𝗮 𝗹𝗼𝗰𝗮𝗹 𝗠𝗖𝗣 𝘀𝗲𝗿𝘃𝗲𝗿 𝗶𝗻 𝘁𝗵𝗲 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗠𝗮𝗰 𝗮𝗽𝗽, and Claude Code can connect to it: Opus 5.5 writes the code and uses it to click through pages and take screenshots.
Codex's computer use is one of the things I use most right now, and it's really fast. I use it all the time for browser testing: open the live site and the local version, compare what's different, then have it fix things. Claude's built-in browser has never worked well for me, and now I can use this from Claude Code too.
It can only control apps you've approved in Codex, and Terminal is blocked outright. It's unofficial, so a ChatGPT update could break it any time, but right now it's ridiculously good.
Send this to Claude Code 👇
"Set up Codex's computer use as an MCP server for you on my Mac:
1. If there's already an MCP server called codex-cu and it works, skip to step 3.
2. Otherwise, open the .mcp.json in the newest version folder under ~/.codex/plugins/cache/openai-bundled/unified-computer-use/ (if it's not there, search ~/.codex for a .mcp.json that contains cua_repl). Copy cua_repl's command, args and env exactly, and register it as codex-cu with claude mcp add-json --scope user. Then stop and tell me to start a new session, and I'll send this again.
3. It can only control apps I've approved in Codex. If you hit "not approved", tell me which app to approve in Codex, and don't switch to another app to get around it. Don't use browser backends like chrome or the in-app browser. Control the browser as a normal app.
4. Test: in a browser I've approved, open a new tab and load [live URL] and then [local URL, e.g. http://localhost:3000]. Take a screenshot of each and list what's different. If a URL is missing or the local server isn't running, ask me first. Don't type passwords or tokens, don't click submit, pay or send buttons, don't touch tabs I already have open, and close the tab you opened when you're done.
5. Once I confirm the differences, fix the local code and compare screenshots again."
*taps sign*
I think Claude Code is overall a great product, i loved using it for the past 2 weeks.
Only two things where they brutally lose to Codex are:
- Computer Use. I think everyone knows that, but Codex’s computer use is by far SOTA, better, faster, more capable than any other on the market, especially claude’s which is very bad
- Remote control. It looks good, and works good, especially from the ios app, but u cant just make a new session in every directory just when claude app is open. U have to use “claude rc” in every dir u want to work in. Thats shitty ux.
I hope someone from Anthropic sees this, and maybe fixes those two. I’d really love them to, it would make it a product so much better.