Meta now lets Claude run your ad account from one chat. Lock the switch that spends your money before you connect it.
It's Meta's own Meta Ads MCP. Claude builds campaigns, writes ads, sets audiences and reads every number.
Everything it creates lands paused. One tool turns it on, ads_activate_entity, and Meta's docs say it starts spending budget.
So I blocked it in Claude Code:
1. Add the Meta Ads MCP with claude mcp add
2. In .claude/settings.json deny mcp__meta-ads__ads_activate_entity and mcp__meta-ads__ads_boost_ig_post
3. Put mcp__meta-ads__ads_update_entity under ask
First question I asked: which ads spent money last month and got nothing back.
6 ads, $412, gone. Claude rebuilt them as 12 drafts and couldn't launch one. I turned on 3 myself.
Do this before you let any agent near your ad budget.
18
63
1,195
Claude Code wrote me a perfect cold email, then admitted it had no idea who to send it to.
It can't look up a real work email. It guesses, and guessed emails bounce.
So I connected Wiza MCP. That's Wiza's 800M+ contacts right inside the chat, and setup is one command in the terminal and a sign-in.
Then I just ask. 5 prompts I now reuse every week:
1. Leads: "Find SaaS companies in Austin with 51-200 people, then their VP Sales. Save the top 50 as a list with work emails and phone numbers."
2. Hiring: "Find senior iOS engineers at fintech companies in Berlin. Save 30 with work emails."
3. After an event: "Here are 40 names and companies from the conference. Make a Wiza list and find their emails."
4. Marketing list: "Find heads of marketing at UK skincare brands with 11-50 people."
5. Before sending: "Show me only contacts with valid emails."
The first one took 6 minutes. I got 50 names, titles and work emails, and Claude wrote a first line for each from their company info.
It also works in Claude Desktop, ChatGPT and Codex. The free plan gives you 20 emails and 5 phone numbers to test it.
Setup and the free plan: wiza.co/mcp
Next time Claude asks who the email is for, let it look them up.
21
59
870
I gave Claude Code one job: change one word in my project. It burned 14% of my 5-hour limit doing it.
By default Claude works one step at a time. It opens a file, reads it, comes back, then does the next one. 38 files, 77 trips, and every trip eats your limit.
On November 12 Anthropic added Code Mode. Claude writes one small script that does all the steps at once and reads back only the part it needs.
I ran the same job twice on Opus 5.5:
Default: 77 trips, 9 min, 14% of my 5-hour limit
Code Mode: 3 trips, 3 min, 5% of my 5-hour limit
Anthropic didn't publish a single number for this, so these are mine.
How to turn it on:
1. Update Claude Code to 2.1.323 or newer
2. Add this to ~/.claude/settings.json:
{ "env": { "CLAUDE_CODE_EXPERIMENTAL_CODE_MODE": "1" } }
3. Restart and pick "Yes, turn on Code Mode"
It pays off on 3 kinds of jobs:
→ the same change across many files
→ a big search you want filtered before Claude reads it
→ several commands that can run at once
It only works in the terminal for now, not in the Desktop app or VS Code. It's experimental, so it can change in any update.
Turn it on before your next big cleanup, not after the limit warning.
I gave Claude Code one job: change one word in my project. It burned 14% of my 5-hour limit doing it.
By default Claude works one step at a time. It opens a file, reads it, comes back, then does the next one. 38 files, 77 trips, and every trip eats your limit.
On November 12 Anthropic added Code Mode. Claude writes one small script that does all the steps at once and reads back only the part it needs.
I ran the same job twice on Opus 5.5:
Default: 77 trips, 9 min, 14% of my 5-hour limit
Code Mode: 3 trips, 3 min, 5% of my 5-hour limit
Anthropic didn't publish a single number for this, so these are mine.
How to turn it on:
1. Update Claude Code to 2.1.323 or newer
2. Add this to ~/.claude/settings.json:
{ "env": { "CLAUDE_CODE_EXPERIMENTAL_CODE_MODE": "1" } }
3. Restart and pick "Yes, turn on Code Mode"
It pays off on 3 kinds of jobs:
→ the same change across many files
→ a big search you want filtered before Claude reads it
→ several commands that can run at once
It only works in the terminal for now, not in the Desktop app or VS Code. It's experimental, so it can change in any update.
Turn it on before your next big cleanup, not after the limit warning.
Anthropic took $540 from me for usage I never touched, and the proof sat on my Claude Usage page the whole time.
The math is on Claude's own pricing page:
Pro, $20: 1x usage
Max, $100: 5x Pro
Ultra, $200: 10x Pro a week
So the weekly bar on a big plan is really a bar on the small one.
On Ultra:
under 10% a week → Pro covers you, you save $180 a month
under 50% → Max covers you, you save $100
On Max:
under 20% a week → Pro covers you, you save $80
My Ultra bar never went past 8%. That's $540 I paid for a meter I never filled.
Check yours in 1 minute:
1. Open Settings > Usage
2. Look at the weekly bar the day before it resets
3. Do it for 2 weeks and take the higher week
Two catches: Pro keeps a 5-hour limit, and Fable on Pro runs on paid credits. If you live in Fable or work in long bursts, Max is your floor.
Do it before your next billing date.
Anthropic took $540 from me for usage I never touched, and the proof sat on my Claude Usage page the whole time.
The math is on Claude's own pricing page:
Pro, $20: 1x usage
Max, $100: 5x Pro
Ultra, $200: 10x Pro a week
So the weekly bar on a big plan is really a bar on the small one.
On Ultra:
under 10% a week → Pro covers you, you save $180 a month
under 50% → Max covers you, you save $100
On Max:
under 20% a week → Pro covers you, you save $80
My Ultra bar never went past 8%. That's $540 I paid for a meter I never filled.
Check yours in 1 minute:
1. Open Settings > Usage
2. Look at the weekly bar the day before it resets
3. Do it for 2 weeks and take the higher week
Two catches: Pro keeps a 5-hour limit, and Fable on Pro runs on paid credits. If you live in Fable or work in long bursts, Max is your floor.
Do it before your next billing date.
Elon Musk's team just made the project manager useless. In Grok Bot 1.0, your bots hand work to each other, and you only see what's finished.
Bots can now message each other. The one that gets the message wakes up, does the work and replies later.
I'm out of the middle.
So I rebuilt my setup around one primary Bot, the one with the star in the sidebar. It checks in with me on its own and hands work to the others.
My team:
- Chief (primary): takes my request and picks who does what
- Researcher: finds sources and links every claim
- Writer: turns the notes into a draft
- Reviewer: checks the draft against the sources
On Monday I sent Chief one line: "a newsletter about my 3 best clients this year, by Friday."
- 09:04, Chief splits it into 3 jobs and messages Researcher
- 11:30, Researcher hands 14 sourced notes to Writer
- 15:10, Writer sends the draft to Reviewer
- 15:40, Reviewer flags 2 blocking issues and Writer fixes them
- Tuesday 08:00, Chief drops the finished draft in my chat
I opened 1 chat that week instead of 12.
Paste this into your primary Bot:
"You run my other Bots. Give every stage one owner and never put two Bots on the same step. Reviewer checks every draft against the sources and lists only blocking issues. Bring me only what passed review. Don't publish or send anything without my yes."
Send it to your Grok Bot before Monday.
Elon Musk's team just made the project manager useless. In Grok Bot 1.0, your bots hand work to each other, and you only see what's finished.
Bots can now message each other. The one that gets the message wakes up, does the work and replies later.
I'm out of the middle.
So I rebuilt my setup around one primary Bot, the one with the star in the sidebar. It checks in with me on its own and hands work to the others.
My team:
- Chief (primary): takes my request and picks who does what
- Researcher: finds sources and links every claim
- Writer: turns the notes into a draft
- Reviewer: checks the draft against the sources
On Monday I sent Chief one line: "a newsletter about my 3 best clients this year, by Friday."
- 09:04, Chief splits it into 3 jobs and messages Researcher
- 11:30, Researcher hands 14 sourced notes to Writer
- 15:10, Writer sends the draft to Reviewer
- 15:40, Reviewer flags 2 blocking issues and Writer fixes them
- Tuesday 08:00, Chief drops the finished draft in my chat
I opened 1 chat that week instead of 12.
Paste this into your primary Bot:
"You run my other Bots. Give every stage one owner and never put two Bots on the same step. Reviewer checks every draft against the sources and lists only blocking issues. Bring me only what passed review. Don't publish or send anything without my yes."
Send it to your Grok Bot before Monday.
Fox retweeted
BUILD AN AI TRADING RESEARCH TEAM FOR FREE. TURN ITS IDEA INTO A STRATEGY YOU CAN TEST.
Start with one thesis: AI infrastructure demand keeps growing.
Now give an AI research team four jobs:
→ SCOUT: find public companies and ETFs exposed to AI infrastructure
→ ANALYST: compare their business exposure and risks
→ STRATEGIST: turn the thesis into entry and exit rules
→ SKEPTIC: challenge the assumptions and identify what could break the idea
Take the resulting strategy and paste it into @horizon_trade_x
Horizon helps structure the rules and backtest them on historical data before you decide whether to paper trade or connect a supported broker.
THE WORKFLOW:
AI THESIS → RESEARCH → TRADING RULES → BACKTEST → YOUR DECISION
You can start free and try Horizon’s AI with your own idea. No coding required.
The prompt and the test come before any real money moves.
The full information is below. Save it.
I turned off my Fable advisor in Claude Code after one paragraph in Anthropic's own guide. 2.1x the money for 1.7 points, and they call it noise.
The advisor setup is all over my feed this week. A cheaper model writes the code, Fable wakes up to think at the big moments.
Anthropic tested that pairing: Opus 5.5 doing the work, a Fable 5.1 advisor on top.
90.1% with the advisor
88.4% without it
2.1x the cost per attempt
Their words: "at the edge of run-to-run noise."
If you set it up this week, /advisor off turns it off.
The same guide has a trick that went the other way. It's cheaper and it scores higher:
Opus 5.5 on high for every task: 95.3%, $0.29 a task
Opus 5.5 on low, failures rerun on high: 97%, $0.17 a task
It only works when something can check the result for you, like a test or a build.
When Claude Code isn't good enough, Anthropic's order is:
1. Give it a way to check its own work
2. Raise effort one level
3. Only then move up a model, and start it at a lower effort
And go easy on xhigh. On Opus 5.5 it bought 1.4 points for 2.5x the price of high.
Save this for the day your weekly limit dies on a Tuesday.
You've been paying Claude Opus 5.5 senior-engineer money to find one number in a PDF.
XERJ Reader launched today, and it does that job over 1,000x cheaper.
It's a different kind of model. It doesn't write the answer, it points at it: you ask in plain words, it highlights the exact line in your file.
A million questions on a short document:
GPT-6 Astra: $17.2K
Claude Opus 5.5: $6.9K
Claude Haiku 5.5: $172
XERJ Reader: $6.50
The part I like more than the price: it can't invent a number. I asked it for the monthly fee in a contract that doesn't have one, and it answered "not in the document".
Where it pays off:
- invoices: total, due date, PO number
- contracts: notice periods and auto-renewal traps
- emails: who, when, how much
- logs and CSVs nobody wants to read
One answer takes 80 ms. You pay only for what it reads, the answer itself is free.
The waitlist opened today. Next time your agent spends a dollar reading a PDF, give that job to Reader.
I turned off my Fable advisor in Claude Code after one paragraph in Anthropic's own guide. 2.1x the money for 1.7 points, and they call it noise.
The advisor setup is all over my feed this week. A cheaper model writes the code, Fable wakes up to think at the big moments.
Anthropic tested that pairing: Opus 5.5 doing the work, a Fable 5.1 advisor on top.
90.1% with the advisor
88.4% without it
2.1x the cost per attempt
Their words: "at the edge of run-to-run noise."
If you set it up this week, /advisor off turns it off.
The same guide has a trick that went the other way. It's cheaper and it scores higher:
Opus 5.5 on high for every task: 95.3%, $0.29 a task
Opus 5.5 on low, failures rerun on high: 97%, $0.17 a task
It only works when something can check the result for you, like a test or a build.
When Claude Code isn't good enough, Anthropic's order is:
1. Give it a way to check its own work
2. Raise effort one level
3. Only then move up a model, and start it at a lower effort
And go easy on xhigh. On Opus 5.5 it bought 1.4 points for 2.5x the price of high.
Save this for the day your weekly limit dies on a Tuesday.
You've been paying Claude Opus 5.5 senior-engineer money to find one number in a PDF.
XERJ Reader launched today, and it does that job over 1,000x cheaper.
It's a different kind of model. It doesn't write the answer, it points at it: you ask in plain words, it highlights the exact line in your file.
A million questions on a short document:
GPT-6 Astra: $17.2K
Claude Opus 5.5: $6.9K
Claude Haiku 5.5: $172
XERJ Reader: $6.50
The part I like more than the price: it can't invent a number. I asked it for the monthly fee in a contract that doesn't have one, and it answered "not in the document".
Where it pays off:
- invoices: total, due date, PO number
- contracts: notice periods and auto-renewal traps
- emails: who, when, how much
- logs and CSVs nobody wants to read
One answer takes 80 ms. You pay only for what it reads, the answer itself is free.
The waitlist opened today. Next time your agent spends a dollar reading a PDF, give that job to Reader.
You've been paying Claude Opus 5.5 senior-engineer money to find one number in a PDF.
XERJ Reader launched today, and it does that job over 1,000x cheaper.
It's a different kind of model. It doesn't write the answer, it points at it: you ask in plain words, it highlights the exact line in your file.
A million questions on a short document:
GPT-6 Astra: $17.2K
Claude Opus 5.5: $6.9K
Claude Haiku 5.5: $172
XERJ Reader: $6.50
The part I like more than the price: it can't invent a number. I asked it for the monthly fee in a contract that doesn't have one, and it answered "not in the document".
Where it pays off:
- invoices: total, due date, PO number
- contracts: notice periods and auto-renewal traps
- emails: who, when, how much
- logs and CSVs nobody wants to read
One answer takes 80 ms. You pay only for what it reads, the answer itself is free.
The waitlist opened today. Next time your agent spends a dollar reading a PDF, give that job to Reader.
🤝 Paid partnership
I've built a new AI model that is not allowed to generate text. Instead, it produces a verifiable program that answers your prompt using the original reference.
XERJ Reader-1 is a neurosymbolic model for working with data. It points to the exact words in your document, writes a small program over them, and exact code computes the answer:
• 20-90x faster
• 10-1,000x cheaper (output tokens free)
• every answer is a program you can replay
The shortest path to automating data work.
This is f*cking insane. Boris Cherny, the guy who built Claude Code, runs 5 Claudes at once. You're still waiting for one to finish typing.
And it just got cheaper. Since Nov 4, background agents in Claude Code use about 5% less of your limit.
Last night I gave Claude Code 5 jobs in one message:
sort 3 years of receipts into a spreadsheet
rewrite my landing page
find broken links on my site
turn a 2-hour call into notes
rename 400 photos by date
One by one: 2h 50m.
All at once: 41 minutes.
How:
1. Update Claude Code to v2.1.316 or newer
2. Put all the jobs in one message and add "run each one as a background agent"
3. Ctrl+B sends a running task to the background
4. /tasks shows who's still working
Give each agent its own files, or they'll step on each other.
Next time you're watching Claude type, that's 4 jobs you didn't start.
Content creators are getting f*cked by their own routine
15 hours a week on work about work
10 of them never needed a human
Here are 10 repos that run the whole content desk for $0:
1. Wallabag → saves every link so the source desk reads 12 and keeps 4
github.com/wallabag/wallabag
2. RSS-Bridge → turns sites with no feed into a feed
github.com/RSS-Bridge/rss-br…
3. yt-dlp → pulls the raw footage and references in one line
github.com/yt-dlp/yt-dlp
4. auto-editor → cuts the silence out of 10 raw videos while you sleep
github.com/WyattBlue/auto-ed…
5. editly → builds clips from a plain config, no timeline
github.com/mifi/editly
6. Subtitle Edit → captions fixed and synced in minutes
github.com/SubtitleEdit/subt…
7. tldraw → 3 thumbnail sketches before you open a design app
github.com/tldraw/tldraw
8. Ghost → the newsletter and the site in one place
github.com/TryGhost/Ghost
9. DocuSeal → sponsor contracts signed without a paid plan
github.com/docusealco/docuse…
10. Matomo → 3 observations, 2 hypotheses, 1 experiment from your own numbers
github.com/matomo-org/matomo
Put one bot on top with 2 routines:
> weekday digest at 09:00
> Friday review at 16:00
> nothing gets posted until you say yes
5 runs a week instead of 168
1 newsletter, 3 clips, 5 posts from a single idea
The stack you pay for today costs $432 a year
This one costs an evening of setup
Save it before your next 15 hour week starts
This is f*cking insane. Boris Cherny, the guy who built Claude Code, runs 5 Claudes at once. You're still waiting for one to finish typing.
And it just got cheaper. Since Nov 4, background agents in Claude Code use about 5% less of your limit.
Last night I gave Claude Code 5 jobs in one message:
sort 3 years of receipts into a spreadsheet
rewrite my landing page
find broken links on my site
turn a 2-hour call into notes
rename 400 photos by date
One by one: 2h 50m.
All at once: 41 minutes.
How:
1. Update Claude Code to v2.1.316 or newer
2. Put all the jobs in one message and add "run each one as a background agent"
3. Ctrl+B sends a running task to the background
4. /tasks shows who's still working
Give each agent its own files, or they'll step on each other.
Next time you're watching Claude type, that's 4 jobs you didn't start.
Content creators are getting f*cked by their own routine
15 hours a week on work about work
10 of them never needed a human
Here are 10 repos that run the whole content desk for $0:
1. Wallabag → saves every link so the source desk reads 12 and keeps 4
github.com/wallabag/wallabag
2. RSS-Bridge → turns sites with no feed into a feed
github.com/RSS-Bridge/rss-br…
3. yt-dlp → pulls the raw footage and references in one line
github.com/yt-dlp/yt-dlp
4. auto-editor → cuts the silence out of 10 raw videos while you sleep
github.com/WyattBlue/auto-ed…
5. editly → builds clips from a plain config, no timeline
github.com/mifi/editly
6. Subtitle Edit → captions fixed and synced in minutes
github.com/SubtitleEdit/subt…
7. tldraw → 3 thumbnail sketches before you open a design app
github.com/tldraw/tldraw
8. Ghost → the newsletter and the site in one place
github.com/TryGhost/Ghost
9. DocuSeal → sponsor contracts signed without a paid plan
github.com/docusealco/docuse…
10. Matomo → 3 observations, 2 hypotheses, 1 experiment from your own numbers
github.com/matomo-org/matomo
Put one bot on top with 2 routines:
> weekday digest at 09:00
> Friday review at 16:00
> nothing gets posted until you say yes
5 runs a week instead of 168
1 newsletter, 3 clips, 5 posts from a single idea
The stack you pay for today costs $432 a year
This one costs an evening of setup
Save it before your next 15 hour week starts
A year ago AI could do 2.5% of real Upwork jobs well enough to get paid. Claude Fable 5.2 just did 41%.
Test is called the Remote Labor Index. 358 Upwork freelancers handed over real paid projects: games, product design, architecture, video, data work.
The average job took a human 29 hours and paid $633. The AI had a $30 generation budget.
Trained human judges scored every result by one rule: would a reasonable client accept this?
Oct 2025, best AI: 2.5%
GPT-6 Astra: 20.8%
Opus 5.5: 21.3%
Claude Fable 5.2: 41.3%
The test left out 3 kinds of work: talking to clients, physical labor, and results that take months to show.
That's where I'd be moving my income this year.
Paste this into Claude with your job title and your last 20 tasks:
"Sort each task into one of 3 lists: AI can deliver this today at a level a client would accept, AI makes me 2x faster at it, or it still needs me. Be harsh. Then tell me what to sell more of next year."
Run it before you send your next invoice.
A year ago AI could do 2.5% of real Upwork jobs well enough to get paid. Claude Fable 5.2 just did 41%.
Test is called the Remote Labor Index. 358 Upwork freelancers handed over real paid projects: games, product design, architecture, video, data work.
The average job took a human 29 hours and paid $633. The AI had a $30 generation budget.
Trained human judges scored every result by one rule: would a reasonable client accept this?
Oct 2025, best AI: 2.5%
GPT-6 Astra: 20.8%
Opus 5.5: 21.3%
Claude Fable 5.2: 41.3%
The test left out 3 kinds of work: talking to clients, physical labor, and results that take months to show.
That's where I'd be moving my income this year.
Paste this into Claude with your job title and your last 20 tasks:
"Sort each task into one of 3 lists: AI can deliver this today at a level a client would accept, AI makes me 2x faster at it, or it still needs me. Be harsh. Then tell me what to sell more of next year."
Run it before you send your next invoice.
Fox retweeted
COINBASE’S CEO IS RUNNING AN AI TEAM THAT MANAGES ITS OWN AI TEAMS.
One agent oversees the operation.
Lead agents manage different areas. Each lead directs sub-agents that write code, test changes and coordinate reviews.
Brian Armstrong doesn’t have to jump between dozens of chats.
Everything that needs his attention goes into one prioritized review queue.
Here’s the architecture:
→ AI Chief of Staff: oversees the operation
→ Lead agents: manage separate workstreams
→ Sub-agents: build, test and coordinate
→ Review queue: delivers decisions, screenshots and video demos
→ Human: reviews the work and makes the final call
There’s another layer.
The system learns Brian’s preferences and distills them into what he calls a "Brian brain"
The goal: better recommendations and fewer corrections before the work reaches him.
Now imagine applying this to a Grok Bot stack:
RESEARCH LEAD scans papers, launches and GitHub MARKET LEAD tracks prediction markets and signals CONTENT LEAD turns findings into post ideas
Each lead runs its own workflow. The Chief of Staff collects the strongest results and sends them to one queue.
One person. An AI organization. One place to make decisions.
A founder with 40 followers on X has $1.3M in revenue. You have 10x the followers, a ChatGPT business plan and $0.
I asked ChatGPT for 10 apps a vibe coder could build, and it called every one of them huge.
Claude Code then checked them on TrustMRR, the site that reads revenue straight from founders' payment accounts.
Only 2 of the 10 had real money behind them. The astrology app made $46 and the air fryer app made $88.
What people paid for in the last 30 days is painfully boring:
• Comp AI, security compliance paperwork: $898K
• Cometly, tracking which ads make money: $212K
• Rezi, a resume builder: $208K
• Kibu, software for disability care providers: $200K
• bytemail, a $19 Mac email app from one person: $23K
• BasinCheck, safety audits for oil and gas contractors: $2.2K
2 out of 3 businesses on TrustMRR sit under $1K.
Send the prompt below to Claude Code before you build your next idea.










