If leads sit in your CRM and nobody owns the next step, that's what I fix for service firms. Agents with a human check. Pinned: my Grok Bots setup.
Joined January 2021
- Tweets4.1K
- Following815
- Followers295
- Likes1.4K
Pinned Tweet
my approval gate said yes to things it had never seen. for weeks
that and 3 other things broke while i set up 15 Grok Bots to run my business. here's the whole setup, every bot, and the prompts i'd paste tonight
7
3
645
Anthropic's own advice on Fable 5.2: start on Opus 5.5, and only move a job when your Opus evals at higher effort still miss the bar.
Write the eval before you switch the model ID.
Fable 5.2 is $8 / $40 per million tokens on Anthropic's published list, twice Opus 5.5 at $4 / $20.
A setup guide from those pages: when to stay on Opus, four Fable 5.1 breaking changes, and the 30-day retention check. No live client job on it yet.
2
1
22
Operators, what's the one task you'd hand to an AI agent tomorrow if you trusted it not to break?
10
Fable 5.2 is $8 / $40 per million tokens on Anthropic's published list, twice Opus 5.5 at $4 / $20.
A setup guide from those pages: when to stay on Opus, four Fable 5.1 breaking changes, and the 30-day retention check. No live client job on it yet.
Sources I opened on 10 Nov 2026 (Anthropic's published list, not my costs):
Announcement: anthropic.com/claude-fable-a…
Overview: platform.claude.com/docs/en/…
What's new: platform.claude.com/docs/en/…
Pricing: platform.claude.com/docs/en/…
Promo note: support.claude.com/en/articl…
7
Rule I keep on every bot that drafts outbound mail:
Anything that spends or sends to a customer waits for me.
The bot can draft.
The bot can queue.
I still press send after I read it.
1
18
A local open source studio for images, video, music, voice and sfx, plugged straight into Hermes. The plugin part is what matters to me.
I already use Hermes to post exact text I write. The step that still breaks is the visual. If the agent can call the studio itself, that handoff goes away.
Imagination Studio... The Hermes Plugin:
Our all in one local open source creative studio for images,movies,music,voice,motion graphics, sfx + more is Now Available with direct Hermes integration via Hermes Plugin. Our official submission to the Hermes Agent Plugin Hackathon. @u1tra_instinct and me!
How to stop leads going cold between two tools, with a nightly check that only flags.
The most expensive lead in a small agency is the one that filled in the form and never reached the CRM.
How it works:
1. Every night one script compares the form log to the CRM
2. Anything in the form log and not in the CRM gets flagged to the owner
3. Nothing gets sent, nothing gets edited
4. The owner decides in the morning
Boring, and it catches the leak you never see.
13
Dots can now control many machines on top of its own cloud computer.
My one line reply under that launch did 22,120 views.
The other 133 replies from yesterday did 5,929 combined.
291 followers today. So I'm writing fewer questions and more real takes on fresh launches.
1
1
28
A lead fills the form at 9pm and hears back the next afternoon.
A signed client waits on a kickoff nobody booked.
The Friday report says 0 and nobody asks why.
Before you hire for any of these, check who owns the step between the two tools.
Usually nobody does.
22
the demo is a one paragraph prompt turning into a working device.
in a service firm the hard part starts after that. who checks the output, who talks to the client, and who owns it when the handoff drops.
During pre-release testing, Claude Fable 5.2 and Claude Mythos 5.2 worked on ambitious projects in many different fields. Here are a few.
From a one-paragraph prompt, Fable 5.2 designed three working electronic devices, including this home voice assistant hub.
12
If I had to rewrite a vague agent proposal today, I would force countable scope.
Not "improve reporting".
Instead: three report types, two data sources, one owner per field, one done check per report.
Vague scope creates silent work. Countable scope creates change orders when it grows.
1
10
The recruiting firm build ends at an AI operating system dashboard. Step one was a careers page scraper.
Score the job you hate most:
[ ] one workflow
[ ] a done state you can count
[ ] a human presses send
3 of 3 goes first. Anything touching money or customer email waits.
1
52
279 followers. 3,841 posts.
most of those posts are replies that landed hours late.
starting today: reply early, post real numbers, answer everyone who replies.
i'll post the count every week, good or bad.
what moved your first 1k?
34
On an outbound call, the question was when to plug in Lemlist.
I said not yet. Leads were in Clay and in the CRM, some in both.
First, pick which system owns each email, then dedupe.
Next, make retries safe, so a failed step can't email a lead twice.
Then Lemlist.
2
1
27
MiniMax M3.1 Flash is free in OpenCode now. do you route the small jobs there and keep the paid models for the rest?
1
44
do the inbox agents book straight onto the client calendar, or is there still a hold step?
The 1-person, $50k/mo lead gen agency is now possible.
> 5-10 clients paying $5k-$10k/mo
> Light retainer + heavy performance
> Pay per meeting booked, closed deals, etc.
> Onboarding agent handles onboarding comms and initial research,
> List-building agent builds lists based on client ICP in the form and onboarding agent's research
> Offer agent concepts multiple offers for approval based on winning cold email offers
> Copy agent trained on top 1% cold emails drafts sequence copy
> Inbox agents manage and route replies AND are synced with client calendars to book prospects right onto it with no human intervention
> Campaign management agent launches, review results, and tests messaging based on metric benchmarks being met / not met
> Founder handles growth, sales, etc.
Sounds crazy right now, but this is possible.
13
56.9% at default effort on CursorBench, up from 49.2%. does the 60% at max effort show up on messy repo work, or mostly on the bench?
Claude Fable 5.2 is now available in Cursor!
At max effort, it's the top model on CursorBench at 60%.
Fable 5.2 scores 56.9% on CursorBench at default effort, up from 49.2% for Fable 5.1.
16
"your agent will talk to other agents at work"
yeah that's the part I'm building for service firms in HK.
the leak isn't the model. it's the handoff after someone replies.
Email and Slack changed how we work. Personal AI agents could be next.
@getserval co-founder @jakeserval: “Your agent will talk to other agents at work.”
But how do you secure agents as cyber risk rises?
9
this is the bit most firms get wrong on AI agents.
air gapping everything sounds safe until sales and ops still need to talk to tools.
the real question is who owns the agent when it breaks a lead follow up.
The myth: We need to turn the internet off. We gotta air gap everything.
The reality: An agent does the same thing a human would do, just better and faster.
@George_Kurtz on No Priors:
2
25