klöss
@kloss_xyz
AI systems architect. prompt shaman.
creative explorer. @psychanon CEO
6.2K Following 76K Followers
fable 5.1 shipped today and the benchmarks are the least interesting part.
9 things to do with it this week:
1. reopen that expensive project: the one you killed over the bill, when the idea was fine. context-heavy jobs now run up to 45% cheaper than they did on Fable 5.
2. run your old skills through to surgically improve: no edits first, just an audit. find out which half of your instructions were babysitting a dumber model previously.
3. audit which thinking setting each workflow needs from their five levels (low/medium/high/xhigh/max): run the same job at the cheapest and the most expensive, compare. most work doesn’t need what you’re paying for.
4. hand it the bugs giving you the most trouble: a hedge fund had a crash nobody solved in four years. it took apart a vendor’s library and found it.
5. give it a real complex project and walk away. 38 hours unattended at @tryramp, and it caught its own bad result, corrected it, plus kept going. don't check if it finished, check where it wandered.
6. redo your agentic routing: one team clocked it at twice the speed of the tier below on half the tokens, which means every routing call you made in July or August could be overpriced.
7. ask for what used to get refused: the cyber filter in Claude Code now fires about 60% less per session, and the bio one fires 85% less on ordinary medical questions. delete your old workarounds.
8. build your own benchmark system, with 10 to 20 tasks. then, score every AI provider and model on it, plus every release. vendor charts are marketing. your system will be the only one that keeps these releases honest to you.
9. rebuild your content pipeline on it: @canva called the writing the standout upgrade over the code. one search company ran it blind and its judges picked the drafts 2:1 over Fable 5.
if you use Claude, something you run daily is on the wrong model right now…
now go find out which.
Show more
We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.
They're the world’s most advanced models for coding and knowledge work.
A SpaceXAI and Grok Bot team member just dropped 8 @bot templates.
For phone calls, posting on X, writing and editing, generating images, coding, networking, tool testing, and parking tickets.
Steal and remix them.
Show more
meet my new grok @bot tinkabot (v0.1.0)! she helps you make high quality grok bot plugins. you can ask her to look at your APIs, create an MCP and/or skills for it, and then wrap that up as a plugin that you can then submit to us for approval.
once it's approved, your plugin will be available for all grok @bot users to use!
let me know if you run into any issues, i've used it on some small toy APIs but would love to see how well it works on real ones
Show more
fable 5.1 shipped today and the benchmarks are the least interesting part.
9 things to do with it this week:
1. reopen that expensive project: the one you killed over the bill, when the idea was fine. context-heavy jobs now run up to 45% cheaper than they did on Fable 5.
2. run your old skills through to surgically improve: no edits first, just an audit. find out which half of your instructions were babysitting a dumber model previously.
3. audit which thinking setting each workflow needs from their five levels (low/medium/high/xhigh/max): run the same job at the cheapest and the most expensive, compare. most work doesn’t need what you’re paying for.
4. hand it the bugs giving you the most trouble: a hedge fund had a crash nobody solved in four years. it took apart a vendor’s library and found it.
5. give it a real complex project and walk away. 38 hours unattended at @tryramp, and it caught its own bad result, corrected it, plus kept going. don't check if it finished, check where it wandered.
6. redo your agentic routing: one team clocked it at twice the speed of the tier below on half the tokens, which means every routing call you made in July or August could be overpriced.
7. ask for what used to get refused: the cyber filter in Claude Code now fires about 60% less per session, and the bio one fires 85% less on ordinary medical questions. delete your old workarounds.
8. build your own benchmark system, with 10 to 20 tasks. then, score every AI provider and model on it, plus every release. vendor charts are marketing. your system will be the only one that keeps these releases honest to you.
9. rebuild your content pipeline on it: @canva called the writing the standout upgrade over the code. one search company ran it blind and its judges picked the drafts 2:1 over Fable 5.
if you use Claude, something you run daily is on the wrong model right now…
now go find out which.
Show more
We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.
They're the world’s most advanced models for coding and knowledge work.
We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.
They're the world’s most advanced models for coding and knowledge work.
everyone's sharing grok @bot templates.
but blindly importing setups will burn usage and get bots fighting over the same job.
I remixed 7 templates into mine without adding any bots
here's how you can do it too:
1: click a grok bot link on X
2: click Add to Grok Bot in web
3: click the Open in Grok Bot popup
4: do not actually add it in app (this is the whole trick)
5: copy and paste all of its instructions, memories, routines, and/or skills into a new doc
6: give this to your chief of staff bot, and ask it based off what it knows about you which parts should you implement, which parts should you remix, and which current bot or new bot this will apply to
7: remix to your liking then decide
you get all the good parts without bloat.
Show more
do you understand what live models unlock?
> movies that can branch for one viewer
> characters that remember you
> ads made for you vs. picked off a shelf
> assembly lines that fail before they fail
> any classroom you want to walk into
this whole list has been theory for years.
now it's real time, and you can change your mind mid prompt too.
we were buying 8 second video frames.
and now we're buying live worlds.
Show more
Today, we are introducing Orbis 1.0, our first Live Model!
Create living worlds and stream them in real time, with persistent memory, interactivity, and physics-grounded generation of unbounded length.
Try it now at
API available via @reactorworld
Dynamic version:
Stable version:
Show more
Dear algorithm,
Please show this post to everyone who’s a psychonaut, AI creative, and crypto degen.
Tell them bookmarking bangers and turning on notifications are optional but appreciated.
Thank you 🍄
Show more
Everyone says they want Jarvis.
But then they show you another fancy UI while you’re stuck at a keyboard or a chat window.
@poetengineer__ is actually building towards it. Look how she writes with AI.
Closest thing to Tony Stark gestural editing with Jarvis I’ve seen someone build. You stop describing the change and just make it.
Most of the AI industry optimizes the models and leaves the input alone. She skipped that assumption entirely.
And her clips go viral over and over, a lot of the times without anyone giving her credit, which is how you know they’re that good.
Follow her to be a year early on this.
Also, who else is building real agent experience (AX) interfaces?
Let’s get you some spotlight.
Show more
imagining a few small gestures to edit/write with ai:
word: spin synonyms
sentence: shift the tone (cold <-> warm)
paragraph: shorten/expand
A SpaceXAI engineer just shared how he runs his Grok Bot fleet of engineers.
From 15 Cursor cloud agents to 200, with a 30 minute PR loop, 3am audits, and a P0 check every 5 minutes.
Steal or remix his setup.
Show more
had Grok @Bot study 300+ GitHub repos of skills
here are the 12 that influenced my setup:
/humanizer by @blader: make your writing and rewrites sound human
/unslop by @poteto: removes common AI slop patterns from your writing
/no-ai-slop by @petergyang: make your writing more precise and clearly human
/agent-skills by @addyosmani: agents that interview you before they invent any specs
/skills by @mattpocockuk: a suite of S tier engineering and productivity skills
/skills by @MengTo: prompt S tier design and UIs like a design system
/skills by @emilkowalski: design and motion skills that survive real audits
/cartographer by @KingBootoshi: a repo map before your agent wanders
/superpowers by Jesse Vincent: stop guessing when you debug, agentic skills framework + complete software dev methodology forcing agents to follow professional workflows
/gstack by @garrytan: memory improvements + spec, review, and QA playbooks so agents cannot silently edit
/impeccable by @pbakaus: detect, fix, and recheck with 61 rules
/last30days by @mvanhorn: deep research what people said about a topic over the last month
save these and remix them with your Bot!
Show more
A Grok Bot engineer just shared the best article I’ve read this year on agents.
2,000 PRs a month, and all of it rests on self-verifying agents: /create-verification-skill
The takeaway: give agents tools, not markdown.
Most won’t read this.
Make sure you do.
Show more
I'm writing a guide to pstack! Here's part one.
we shipped a lot of improvements to grok @bot last week! quick recap:
one of the simplest paths to generational wealth is to become an attention machine in high value niches...
e.g. AI, crypto, prediction markets
you must own as much digital real estate as possible - YT, X, IG, TikTok, blogs, newsletters
understanding what these audiences want to see & having a system for building brands in them is a serious advantage right now
attention is an asset that you can trade
Show more
my newest Grok Bot (trader) has a phantom wallet, a fomo login, and zero trades to its name
what coin should we buy the top of first?
found another cost saving tip:
Cursor cloud agents can silently flip to fast and up your Grok @bot usage bill 2x.
check which model it's set to in the follow-up field before you send.
it happens after the first reply, even if your default says otherwise.
agents launched from Grok bot bill as cursor-grok-4.6-high-fast when it flips.
this happened to me because my silly ass disabled most models, so it fell back to what was left.
reported on their forums.
Cursor says they're tracking a fix.
until then, make sure the models you want enabled are in your cloud agent app settings.
torched 1B tokens in 7 days lol.
check yours and tell me what you find.
Show more
took this advice and audited my Grok @bot setup as I was mega burning usage
most of mine was testing, setup, and experiments… here were my problems
and how I hopefully fixed them:
→ accidentally had a routine running every 15 minutes… that is 96 times a day (audit your routines)
→ two of the same routine were also running on two bots at once… ngl I still don't know how it happened or why (check across bots, not just within one)
→ a bot I had told to be inactive was still checking websites 12 times a day. telling the bot it was inactive did nothing (once again routines can be separate from a bot… hopefully you're getting the point now)
→ my main chat also got so long that every routine attached to it re-read the whole chat before answering (double check your routines are on the bots that do the job, not all in the chat you talk in)
→ I posted in group channels without tagging a bot… all three responded and answered the same question (tag one bot or pay 3x for the same answer)
→ make sure your chief of staff doesn't confirm a task, chase it, then confirm the receipt… that's four messages for one job (a huge no no… send it once, step away, and have some patience, tweak as you go if you have issues)
→ I ran 800 skill evaluation tests, a lot of them back to back and had bots read the whole skill file each time… fine for a one time test, expensive af when the chain is long (batch them over time, and don't make every bot reread the whole file to check one thing)
→ I had bots forwarding messages to each other across channels. same images and docs sent twice (avoid this setup it's more costly)
→ I was starting more tasks the second a job landed (do one at a time, leave gaps between them, don't auto chain everything it will burn your usage faster)
→ I kept having a bot run a long looping task inside chat instead of running it once (have a Cursor cloud agent write the script, then the bot just runs the file)
my guess is by next week my system could run 60-70% more cost effective and smoother after these changes.
I will post my thoughts either way.
$1,048 is what bad habits can cost us.
save these tips so they don't cost you.
Show more
took this advice and audited my Grok @bot setup as I was mega burning usage
most of mine was testing, setup, and experiments… here were my problems
and how I hopefully fixed them:
→ accidentally had a routine running every 15 minutes… that is 96 times a day (audit your routines)
→ two of the same routine were also running on two bots at once… ngl I still don't know how it happened or why (check across bots, not just within one)
→ a bot I had told to be inactive was still checking websites 12 times a day. telling the bot it was inactive did nothing (once again routines can be separate from a bot… hopefully you're getting the point now)
→ my main chat also got so long that every routine attached to it re-read the whole chat before answering (double check your routines are on the bots that do the job, not all in the chat you talk in)
→ I posted in group channels without tagging a bot… all three responded and answered the same question (tag one bot or pay 3x for the same answer)
→ make sure your chief of staff doesn't confirm a task, chase it, then confirm the receipt… that's four messages for one job (a huge no no… send it once, step away, and have some patience, tweak as you go if you have issues)
→ I ran 800 skill evaluation tests, a lot of them back to back and had bots read the whole skill file each time… fine for a one time test, expensive af when the chain is long (batch them over time, and don't make every bot reread the whole file to check one thing)
→ I had bots forwarding messages to each other across channels. same images and docs sent twice (avoid this setup it's more costly)
→ I was starting more tasks the second a job landed (do one at a time, leave gaps between them, don't auto chain everything it will burn your usage faster)
→ I kept having a bot run a long looping task inside chat instead of running it once (have a Cursor cloud agent write the script, then the bot just runs the file)
my guess is by next week my system could run 60-70% more cost effective and smoother after these changes.
I will post my thoughts either way.
$1,048 is what bad habits can cost us.
save these tips so they don't cost you.
Show more
my number 1 tip for controlling your grok @bot usage is to avoid scheduled routines that run too frequently. for example, a 15 min routine runs almost 100 times a day, and every run consumes tokens. hourly or a few times a day is usually good enough.
the length of the chat with your bot can also make routines much more expensive. for recurring ones, try giving that to a fresh bot, while you continue your chat with your main bots (like a chief of staff).
give this post to your bot and ask them to help!
Show more
Invest the next 3 months into Grok Bot.
You could be one of those millionaires.
Prediction: someone will make $1M+ selling @bot templates in the next 3 months
skill-finder was almost one of my new Bots
so was scribe
but then Grok Bot said no
I'd be wasting usage, and bloating my setup
if you got a chief-of-staff, avoid these too:
scheduler
calendar-manager
knowledge-manager
memory-auditor
executive-assistant
project-manager
bookmark-manager
inbox-manager
chief-of-bots
they all absorb into CoS
here are my updated plans:
Show more
here’s my Grok @bot setup so far ↓
chief-of-staff: runs the other four
intel-gatherer: what’s important today
viral-analyst: studying what really works
content-crafter: polishes a rough idea
community-manager: site monitoring
the next 5 bots I'm setting up ↓
art-director: image and video for content
builder: my full stack engineer
trader: paper trading first, live later
clerk: sponsorships, consulting, reports
skill-finder: searches GitHub + 𝕏
will share my full thoughts in an article
Show more
here’s my Grok @bot setup so far ↓
chief-of-staff: runs the other four
intel-gatherer: what’s important today
viral-analyst: studying what really works
content-crafter: polishes a rough idea
community-manager: site monitoring
the next 5 bots I'm setting up ↓
art-director: image and video for content
builder: my full stack engineer
trader: paper trading first, live later
clerk: sponsorships, consulting, reports
skill-finder: searches GitHub + 𝕏
will share my full thoughts in an article
Show more
