Building the steroids for AI @ https://teapawt.org/t.co/IYyOUaF6n2 x https://teapawt.org/t.co/MaGdVluQpG Every site's search data, inside Claude - https://teapawt.org/t.co/AJANypRHzL
Joined January 2024
- Tweets13.4K
- Following309
- Followers1.2K
- Likes2.4K
Pinned Tweet
I'm cancelling all of my Claude subscriptions
($550 in total).
Not because Claude is a terrible model.
Claude is the BEST one out there.
I just feel that I'm always internally forced to max out the limits.
And if I'm not doing that, i feel unproductive.
This really fucks up the creativity side of the business i think.
i'm giving myself 15 more days,
after that shifting to 1x $100 subscription.
and that too will mostly be running my existing loops & crons.
anyone else feel ... "dragged/pulled" by Claude when the limits aren't exhausted yet?
Centralized compute
Centralized toilets
Centralized girlfriends
Centralized cars
Sam, we already have all of that and there’s name for it:
Public toilets, iykyk, public transport
Yet one of the greatest things about life is choice.
That including the choice to privacy and ownership. Of data. Of compute.
sam altman explaining why you shouldn't own your own compute is the most landlord take of the year.
'your ram sits idle 95% of the time.' so does my laptop's battery, my camera and half my house. idle is what owning something looks like.
local models are the one version of ai where nobody can cut your limits overnight, kill your plan, or read your prompts. of course the guy selling the subscription thinks that's a bad allocation
Opus is literally useless for 99% of tasks.
ran the exact same post-writing prompt on opus and sonnet. same inputs, 3 posts each.
all 6 scored 90/100 on our scorecard rated by jev.
sonnet was ~30% faster and costs half.
been paying double out of habit
(Very rare cases where going all opus makes sense because work is complicated)
I DELETED 4,000 LINKEDIN CONNECTIONS. Real people: old colleagues, classmates.
LinkedIn shows a new post to a slice of your own network first. If most of them aren't your audience, they scroll past and the post stalls at a few hundred views. Your network is your first test audience, so it should look like the people you're writing for.
One single Google search has brought 439 clicks to one of my pages.
The search is "linkedin comment generator". My page has averaged spot 4.7 in the results for it since December, and 9% of the people who see it click. Searches like that are small, but every one of them is someone looking for the tool right now.
I want to find and win those searches for 5 software companies, 1-on-1.
Month one starts with:
• Every search your buyers make, matched to one page on your site
• Your pages stuck on Google's page 2 to 5, rewritten first. That's where my fastest wins have come from
• Titles, internal links and broken redirects fixed, and every page checked so Google and Bing index it
Then every month:
• Up to 8 new pages: "alternatives", "X vs Y" and the questions people ask right before they buy
• One free tool page for your niche. One small tool page on my site has brought in 1,740 Google clicks on its own
• What ChatGPT, Perplexity and Google's AI answers say about tools like yours, and fixes to the pages they quote
• Answers in the Reddit and forum threads that rank for your keywords
• A short weekly note: what moved, what didn't, and the next 3 pages I'm on
You give me read access to Search Console and a login to edit the site. WordPress, Webflow or a GitHub repo all work. We talk on Slack or WhatsApp.
$1,000/mo. If none of your pages move up on Google in the first 30 days, I refund month one. The work stays yours either way.
5 spots. DM me by Tuesday.
sam altman explaining why you shouldn't own your own compute is the most landlord take of the year.
'your ram sits idle 95% of the time.' so does my laptop's battery, my camera and half my house. idle is what owning something looks like.
local models are the one version of ai where nobody can cut your limits overnight, kill your plan, or read your prompts. of course the guy selling the subscription thinks that's a bad allocation
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
Apps that nag you all day are the worst, so I gave mine strict rules 😤
Safe to Quit sends at most 3 nudges a day and never between 10 PM and 8 AM. Tap "Keep it" and it won't ask about that app again today. It also learns what's normal, so a video app that's always busy doesn't nag you.
chatgpt rewrites your buyer's question into its own google-style searches before it reads a single page. nectiv ran ~4,000 prompts through gpt-5.6: 7.61 searches per prompt on average, 10.7 for software
you can read every one of those searches. 3 ways, 10 min
> 1. the free way (browser)
- ask something that forces a search: "best [your category] for [your buyer] 2026"
- devtools → network, then reload the chat
- filter by the conversation id (the bit after /c/ in the URL)
- open the conversation response, ctrl+f search_model_queries (or just "queries")
- heads up: openai hid this field for a stretch mid-2026 and it came back on gpt-5.6. if it's missing, use #2
> 2. the reliable way (api)
- openai responses api with the web_search tool
- the searches sit in web_search_call → action.queries
- action.query only holds the first one. parse that alone and a 3-search fan-out looks like 1
- set user_location to your market. leave it out and you get what a US user triggers
- $10 per 1k searches plus tokens. 30 prompts costs pocket change
- the api runs a different system prompt than the app. great for patterns, a bit off on exact wording
> 3. gemini and claude hand theirs over officially
- gemini api + google search grounding: groundingMetadata.webSearchQueries
- claude api web search: each query sits in server_tool_use → input.query
- dataforseo's chatgpt llm responses endpoint returns fan_out_queries (sometimes empty)
> what to do with them
- run 20-30 prompts your buyer would actually type, dump every query in a sheet, count the repeated words
- ahrefs, 15M fan-outs: "price" fan-outs up 47% in one month, "official" in 13.1% of english ones
- 6.8% are site: searches and 76% of the top site: targets are brand sites. put pricing, specs and comparisons on your own site in plain text
- when the query names a platform (g2, checkatrade, reddit), get listed there before you write another blog post
- run each query through google and bing yourself. ahrefs found only 16.6% of chatgpt's results show up in google's results for the same query
- re-run monthly. it shifts with every model
> read it honestly
- free logged-out chatgpt searched far less: 14.4% of responses had fan-outs, ~1.6 each (ahrefs, 66.6M responses)
- thinking mode on a paid account ran 15-40 on saas prompts
- test the way your buyers actually use it
bookmark this and run your top 10 buyer prompts before you write your next page
chatgpt rewrites your buyer's question into its own google-style searches before it reads a single page. nectiv ran ~4,000 prompts through gpt-5.6: 7.61 searches per prompt on average, 10.7 for software
you can read every one of those searches. 3 ways, 10 min
> 1. the free way (browser)
- ask something that forces a search: "best [your category] for [your buyer] 2026"
- devtools → network, then reload the chat
- filter by the conversation id (the bit after /c/ in the URL)
- open the conversation response, ctrl+f search_model_queries (or just "queries")
- heads up: openai hid this field for a stretch mid-2026 and it came back on gpt-5.6. if it's missing, use #2
> 2. the reliable way (api)
- openai responses api with the web_search tool
- the searches sit in web_search_call → action.queries
- action.query only holds the first one. parse that alone and a 3-search fan-out looks like 1
- set user_location to your market. leave it out and you get what a US user triggers
- $10 per 1k searches plus tokens. 30 prompts costs pocket change
- the api runs a different system prompt than the app. great for patterns, a bit off on exact wording
> 3. gemini and claude hand theirs over officially
- gemini api + google search grounding: groundingMetadata.webSearchQueries
- claude api web search: each query sits in server_tool_use → input.query
- dataforseo's chatgpt llm responses endpoint returns fan_out_queries (sometimes empty)
> what to do with them
- run 20-30 prompts your buyer would actually type, dump every query in a sheet, count the repeated words
- ahrefs, 15M fan-outs: "price" fan-outs up 47% in one month, "official" in 13.1% of english ones
- 6.8% are site: searches and 76% of the top site: targets are brand sites. put pricing, specs and comparisons on your own site in plain text
- when the query names a platform (g2, checkatrade, reddit), get listed there before you write another blog post
- run each query through google and bing yourself. ahrefs found only 16.6% of chatgpt's results show up in google's results for the same query
- re-run monthly. it shifts with every model
> read it honestly
- free logged-out chatgpt searched far less: 14.4% of responses had fan-outs, ~1.6 each (ahrefs, 66.6M responses)
- thinking mode on a paid account ran 15-40 on saas prompts
- test the way your buyers actually use it
bookmark this and run your top 10 buyer prompts before you write your next page
It took me:
- 2 months to make $5 with my first product online
- 2 years to get 1,000 X followers,
- 9 months to get to $1K MRR
- 12+ months to cross $100K in lifetime sales
This was while working 17 hours per day, every day of the week including weekends. Only break was 2 days when I had diarrhoea.
Before I started, from the fake fucks’ tales, it seemed as if the journey goes:
you start,
you blink,
and you’re at $10K MRR.
higgsfield promised to open source its product. never did. so comfyui's community started building the open versions themselves, and higgsfield answered with a cease and desist and a 48 hour deadline.
if your product can be rebuilt in a two week community challenge, a lawyer letter won't save it. if it can't, you didn't need the lawyer. comfy keeping the challenge up is the most based thing an ai company has done this month
Last night, @higgsfield sent @ComfyUI a cease and desist (attached)
They want the Open Source Higgsfield challenge taken down within 48 hours.
Their argument is that our use of the Higgsfield name and that using their CEO's likeness is an infringement violation (man, the irony). Claiming that it makes the event look like Higgsfield is associated with it.
This couldn't be farther from the truth, which is that Higgsfield made a promise to open source its product and never did.
For the past two weeks, builders have been recreating popular closed apps as open workflows anyone can run.
So we asked our community to deliver where Higgsfield failed: apps in the open, built by the community, for the community. We've made it clear that the challenge is organized by ComfyUI and not affiliated with or endorsed by Higgsfield.
When we started Comfy, we made the community a promise. Our product stays open, free, and yours. We’ve kept it every day since, and one letter isn’t going to change that.
The challenge stays up. Comfy and open source will win!
granola just priced dictation at zero. every app charging $12/mo purely to turn speech into text is now on a clock.
we've watched this up close building contextli. people never cared about the transcript. they care whether the slack message comes out sounding like a slack message and the email like an email. raw words were always going to be free
“Claude flagged this conversation of yours because you shared the cure for cancer.
It was deemed a safety hazard. And has been shared with relevant authorities.
Please book your next flight with me to get a 100% discount on the flight”.
We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation.
Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch.
Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve.
Read more on our new Safeguards blog:
safeguards.anthropic.com/202…
I don't trust GA4 anymore. 1,292 of my site's 1,704 visits over 28 days came from "direct", with a 32% engagement rate.
Bots, or something broken?
my site is Locul. Sep 7 to Oct 4:
search console: 184 clicks from google
GA4: 232 sessions from google organic
those two roughly agree. the problem is the other 1,292 "direct" sessions, with a 32% engagement rate
is that bot traffic, or is something misfiring? what would you check first
The most-clapped thing I've written on Medium (15,000+ claps) came from watching a senior developer lose 3 hours to DeepSeek.
She typed a prompt, deleted it, typed another. Then she picked up her phone and just talked the bug through. Fixed in 2 minutes.
Typed: "Help debug React useEffect infinite loop."
Spoken: "This useEffect is fetching user data, something in the dependency array makes it loop, I think because we update state inside, and weirdly it only happens when..."
Same model. The spoken version just had all the details she kept deleting.
Jeff Bezos could make an Amazon executive drop everything with a one-character email: "?"
In 2018 the company shipped billions of packages a year, and he was still reading customer complaints sent to his public address.
Everyone remembers the question mark. The part that made it work is easier to miss 🧵
I learned this the slow way. For about six months, Bing refused to index Contextli's site, even though it was doing fine on Google. I removed the site from Bing Webmaster Tools, added it back, and tried everything else I could find. Nothing worked, and I had honestly given up. Then I found out you can open a support ticket. Microsoft replied within 24 to 48 hours and said the issue was resolved.
So about six months of trying things on my own, and the fix was one support ticket. They said it can take two to three weeks before the site shows up in results again, so I'll come back with the traffic numbers.
I'm going to post one of these every day for a while: a famous company move, and what it looked like when I tried it at my size. Follow along if that sounds useful.