Been building a small agent into my own tooling lately and this matches what I run into. The harness layer trips me up the most.
This is f*cking insane. Former Google CEO Eric Schmidt, the man who ran Google for 10 years:
"If you really want to make money, it's actually easy, found an agentic AI company, I mean build an agent to do something
If you build a better agent than anyone else, you can probably mpany and own it yourself"
In just 27 seconds, he explains why the next 1–2 years are the agentic period, and why the window won't stay open
Prompts → Opus 5.5 → Memory → Skills → MCP → A Company You Own Yourself
Save this before it vanishes from your feed
Everything below is free:
Agent architecture
langchain.com/blog?category_…
Claude Code 101
anthropic.skilljar.com/claude-code-101
Claude Code in action
anthropic.skilljar.com/claude-code-in…
Prompt engineering
platform.claude.com/docs/en/build-…
Interactive prompt course
github.com/anthropics/cou…
Claude.md + memory
code.claude.com/docs/en/claude…
Skills
code.claude.com/docs/en/skills
MCP
code.claude.com/docs/en/mcp
Routines
code.claude.com/docs/en/routin…
Claude Code ultimate guide
github.com/FlorianBruniau…
Awesome Claude Code
github.com/hesreallyhim/a…
Anthropic academy
anthropic.skilljar.com
Official Claude Code docs
code.claude.com/docs/en/overvi…
All of this costs $0
95% of people will spend the entire agentic period scrolling threads
The 5% who actually ship agents will own the companies that come out of it
Bookmark it and start with the first link right now
Then read how to build an agent that actually does the job in the "Context, Harness, Loops, Prompts" guide below ↓
0
0
0
Nick CleanCode retweeted
THIS IS F*CKING GOLD PROMT FOR FABLE 5.2...
Most agent systems don’t fail because the model is weak.
they fail because the harness around it is messy.
Fable 5.2 makes the setup much cleaner:
→ plan the work
→ delegate each task to the right agent
→ verify outputs with nly what actually passed
→ log what happened so the next run gets easier to debug
the key idea is simple:
one task → one owner → one done-check
and before anything gets merged, the harness verifies the result instead of trusting the agent summary.
that’s how you get fewer silent failures, cleaner handoffs, and a system that’s much easier to scale.
full setup below ↓
Most agent systems don’t fail because the model is weak.
they fail because the harness around it is messy.
Fable 5.2 makes the setup much cleaner:
→ plan the work
→ delegate each task to the right agent
→ verify outputs with nly what actually passed
→ log what happened so the next run gets easier to debug
the key idea is simple:
one task → one owner → one done-check
and before anything gets merged, the harness verifies the result instead of trusting the agent summary.
that’s how you get fewer silent failures, cleaner handoffs, and a system that’s much easier to scale.
full setup below ↓
4
1
21
668
Nick CleanCode retweeted
This is f*king insane. SpaceXAI engineer (ex-Cursor):
"Stop prompting your agents. I have 1,000+ of them and haven't written a single prompt. My Chief of Staff bot handles everything."
In this 50-minute session, he shows how to set up the team before you scale it
Job → Chief → 90% of the Routine → 1000+ Agents*
Most people give every bot the whole chat history. That's where the bill comes from
Give each bot one job and a short context once, instead of paying for the same re-reading every run
Worth more than most $2,000 agent engineering courses
Bookmark and watch it today
Then read the full setup: "Grok Bot: 6 Bots, 200+ Agents, a Smaller Bill" ↓
"Stop prompting your agents. I have 1,000+ of them and haven't written a single prompt. My Chief of Staff bot handles everything."
In this 50-minute session, he shows how to set up the team before you scale it
Job → Chief → 90% of the Routine → 1000+ Agents*
Most people give every bot the whole chat history. That's where the bill comes from
Give each bot one job and a short context once, instead of paying for the same re-reading every run
Worth more than most $2,000 agent engineering courses
Bookmark and watch it today
Then read the full setup: "Grok Bot: 6 Bots, 200+ Agents, a Smaller Bill" ↓
1
1
29
1,064
Fun to watch! A family sink has to be one of the most unpredictable environments for a robot. My 11-year-old's dishes alone would be a challenge 🤣
The dishwasher may finally have competition. 🤖🍽️
This robotic kitchen system doesn’t just rinse a plate.
It grabs it, positions it, scrubs it, rinses it and places it neatly into the rack.
That’s the bigger story:
Robotics is moving from factories into the small, repetitive
And once machines become reliable enough to handle kitchens, laundry, cleaning and home maintenance…
the definition of “housework” could change completely.
Would you trust a robot to handle your entire kitchen? 👇
#Robotics #SmartHome #AI #IoT #5G #NLP #Tech #Technology #innovation #Bigdata #Automation #FutureTech
Cc: @jblefevre60 @Nicochan33 @pierrepinna @RosyCoaching @kalydeoo @Ym78200 @enilev @FrRonconii @CurieuxExplorer @Khulood_Almani @mvollmer1 @JoannMoretti @margaretsiegien @Shi4Tech @anand_narang @Eli_Krumova @HaroldSinnott @bimedotcom @EvanKirstel @Hana_ElSayyed @sallyeaves @SpirosMargaris @ipfconline1 @LaurentAlaus @Analytics_699 @Jeffkagan @AndresVilarino @efipm @FMFrancoise @MHcommunicate @Corix_JC @JolaBurnett @JeromeMONANGE @mary_gambara @drsharwood @smoothsale @sulefati7 @debashis_dutta @dinisguarda @devbhatt72 @terence_mills @BetaMoroney @mdrechsler @MarshaCollier @avrohomg @NevilleGaunt @PamMktgNut @arlenenewbigg
0
0
0
1
Whoa, 1 keV with low-speed compression is a real milestone. Congrats to Greg and the team! The hard part is still net gain and commercialization though. I hope the funding holds out long enough to get there.
More than 1 keV achieved. Over 12 million degrees Celsius. Nearly as hot as the core of the Sun. ☀️
Our LM26 fusion demonstration machine recently reached this plasma temperature, a world-first milestone using practical, low-speed compression.
General Fusion CEO Greg Twinney keV is such a significant achievement and what it means for the future of fusion energy.
Watch the video below. 👇
#FusionEnergy #CleanEnergy #GeneralFusion
0
0
0
3
Nick CleanCode retweeted
Share with schools in your network! This year's #TEWeek26 theme #ShapeOurFuture, shows young people how engineering and tech tackles the real-world issues they care about. Get free resources: eukeducation.org.uk/about-us/our-a…
#EngineeringUK25
#EngineeringUK25
0
1
1
143
Nick CleanCode retweeted
I am begging you to paste this into Opus 5.5 before bed...
it will design your first team of AI agents from scratch, with custom prompts, workflows, and a complete 24-hour launch plan
Save this before it takes over your timeline:
it will design your first team of AI agents from scratch, with custom prompts, workflows, and a complete 24-hour launch plan
Save this before it takes over your timeline:
Claude Opus 5.5 isn't just for chatting.
Builders are using it to run teams of AI agents, write content, review and make real money.
Here's how to build your first team from scratch. x.com/i/article/2104…
19
11
124
11,218
Nick CleanCode retweeted
Lauren Tan’s model setup is basically a traffic controller for AI work.
instead of asking one model to handle everything, she decides first:
what kind of thinking does this task actually need
then routes it accordingly:
→ fast model for the first pass
→ deep model when the oning
→ long-context model when the input gets huge
→ reviewer when accuracy matters
→ merge only after the strongest parts survive critique
so the system optimizes for quality, speed, and cost at the same time.
Lauren’s routing map below ↓
instead of asking one model to handle everything, she decides first:
what kind of thinking does this task actually need
then routes it accordingly:
→ fast model for the first pass
→ deep model when the oning
→ long-context model when the input gets huge
→ reviewer when accuracy matters
→ merge only after the strongest parts survive critique
so the system optimizes for quality, speed, and cost at the same time.
Lauren’s routing map below ↓
5
7
70
2,725
Nick CleanCode retweeted
This might be one of the most impressive low-cost humanoid robots I've seen so far
Meet BRIDGE an open-source humanoid developed by researchers from Carnegie Mellon University and their collaborators.
The robot costs around $1,500 to build yet it can perform backflips dance,
reproduce complex human movements.
Meet BRIDGE an open-source humanoid developed by researchers from Carnegie Mellon University and their collaborators.
The robot costs around $1,500 to build yet it can perform backflips dance,
reproduce complex human movements.
Looks simple but watch closely.
Closing a zipper takes constant small corrections at the fingertips. That was my first thought reading about LingBot-Touch, Robbyant's tactile model for its embodied-native robot brain.
@robbyant_brain
1
2
13
594
Nick CleanCode retweeted
Read Aved Electronics' article on why "stability is power" in humanoid robotics, and how battery system design will help shape the future of the industry.
Every movement, balance correction, and micro-adjustment requires power, with stabilisation alone sometimes accounting for a erall battery consumption.
As humanoid robots move closer to widespread deployment, power system architecture could become just as critical as advances in AI.
Read the full article here: humanoidroboticstechnology.com/articles/stabi…
#humanoids #humanoidrobot #humanoidrobotics #humanoid
Every movement, balance correction, and micro-adjustment requires power, with stabilisation alone sometimes accounting for a erall battery consumption.
As humanoid robots move closer to widespread deployment, power system architecture could become just as critical as advances in AI.
Read the full article here: humanoidroboticstechnology.com/articles/stabi…
#humanoids #humanoidrobot #humanoidrobotics #humanoid
0
1
3
116
Nick CleanCode retweeted
This is f*cking insane. Andrej Karpathy spent 8 years at OpenAI and Tesla
Last week, he poured everything he learned into one free 2-hour lecture
Agents → Harness → Loops → Graphs
Some bootcamps charge $15K and still teach you less than this
You may not have 2 hours w and watch before it's gone from your feed
Then read the guide below and build your first agent harness ↓
Last week, he poured everything he learned into one free 2-hour lecture
Agents → Harness → Loops → Graphs
Some bootcamps charge $15K and still teach you less than this
You may not have 2 hours w and watch before it's gone from your feed
Then read the guide below and build your first agent harness ↓
Harness, loop and graph could be the missing layers between AI models and agents that actually work.
Same model, same prompt, completely different results.
In this article, I show you how to build all three layers. x.com/i/article/2119…
14
47
376
37,962
A huge congrats to the CABC team on a great briefing. With manufacturing shifting globally, building these Canada-ASEAN bridges is a smart move. I hope advanced manufacturing gets real traction at the Manila forum.
This week in Montréal, CABC Executive Director Meghan Pritchard addressed representatives of the Aurora Pacific Economic Zone and Freeport Authority (APECO), the Philippine Embassy, and Canadian business leaders at an investment briefing. The event explored new opportunities in hlighted where Canadian expertise and investment can play a role — particularly in defence, energy, infrastructure, logistics, and advanced manufacturing.
We look forward to building on this momentum at CABC's Canada-ASEAN Business Forum in Manila on November 15, bringing together government and business leaders to advance concrete commercial partnerships.
Thank you to Fasken for hosting, and to APECO, the Philippine Embassy, Asian Consulting Group, and the HEC Montréal Entrepreneurship Hub for a great discussion. And thank you to Ambassador Jose Victor V. Chan-Gonzaga, Gil G. Taway IV, Mon Abrea, Mario C. Tani, Angelo Gabriel Mantuano, Winston Chan, Alexander Joseph Estomo, and Ponora Ang for the insights and the conversation.
#CanadaPhilippines #CanadaASEAN #Trade #Investment #ASEAN
0
0
0
10
Nick CleanCode retweeted
Apeco is seeking Canadian investments in aerospace, defense, renewable energy and advanced manufacturing. Its first investment mission to Canada drew over 120 participants.
Read more on newscentralmedia.com.
#NewsCentralMedia #Business #Investment
Read more on newscentralmedia.com.
#NewsCentralMedia #Business #Investment
0
1
0
5
Nick CleanCode retweeted
i still don't understand why everyone is NOT fixing their Grok Bot usage
it will probably make your bots last the whole week & definitely change your life...
i wrote the best article on exactly what your bots need to stop burning your limit
simply paste this article into grok bot,ting your limit and fix it
this is the EXACT setup i ran on my bots, and my limit hasn't run out mid-week since:
it will probably make your bots last the whole week & definitely change your life...
i wrote the best article on exactly what your bots need to stop burning your limit
simply paste this article into grok bot,ting your limit and fix it
this is the EXACT setup i ran on my bots, and my limit hasn't run out mid-week since:
119
88
1,162
170,653
Nick CleanCode retweeted
China wants humanoid robots welding ships.
SPEEDBOT just introduced Weldroid W1, a 170 cm tall humanoid weighing 60 kg.
It has 27 degrees of freedom, a 10 kg arm payload and uses LiDAR and 3D vision to navigate around large metal structures.
The company plans to test it in a the end of 2026.
Keeping the torch steady and producing consistent welds will be a good test.
SPEEDBOT just introduced Weldroid W1, a 170 cm tall humanoid weighing 60 kg.
It has 27 degrees of freedom, a 10 kg arm payload and uses LiDAR and 3D vision to navigate around large metal structures.
The company plans to test it in a the end of 2026.
Keeping the torch steady and producing consistent welds will be a good test.
1
3
12
1,473
Oh nice, this is a real problem. My AI-built screens absolutely look like five different apps 🤣 Will give this a try this weekend.
Your AI-generated screens shouldn’t look like five different apps.
OpenDesigner is a design-system add-on for builders using AI tools like Claude, ChatGPT, Codex, and Cursor.
It helps you keep an app’s look consistent by turning answers about your product into saved design rts that your code, team, and next AI chat can follow.
Key features:
• Project-aware setup – reads your project first so it doesn’t ask what your files already answer
• Guided design choices – answer up to five short questions to get a first design system, then refine color, type, spacing, or motion
• Accessibility checks – checks text contrast, tap targets, keyboard focus, and reduced motion; errors stop the build
• Multiple export formats – generates DTCG tokens, CSS variables, a Tailwind theme, Figma variables, Swift, and Jetpack Compose
• Existing-tool support – install through Claude Code, copy skills into supported agents, or use ChatGPT Project instructions and knowledge files
It’s open-source (MIT license).
It’s early: the README notes there’s no tagged release yet.
Link in the reply 👇
0
0
0
7
Nick CleanCode retweeted
If you still think AI video is a slot machine, try sending this to Claude Code:
"You're the director on this film. You write the story, the storyboard and every prompt, and I only talk to you. All drawing goes to Codex through codex exec, with $imagegen in the prompt, one image order, and don't move on until each step passes:
1. Story first. No images until I say yes.
2. Have Codex draw one sample frame. Once I pick the style, start on the real shots.
3. For each shot, have Codex draw the first and last frame. Zoom in on every one, and fix mistakes on the image so they never make it into the video.
4. Before you submit any video, spin up a fresh agent that hasn't seen anything and have it only look for mistakes. Once it says PASS, give me the price.
5. Only run the video after I say yes."
All the drawing runs on Codex, 𝗳𝗿𝗲𝗲 with your ChatGPT subscription, on the strongest image engine out right now, 𝗜𝗺𝗮𝗴𝗲𝘀 𝟮.𝟲. The AI shorts you've seen that don't look AI-made were shot exactly like this.
"You're the director on this film. You write the story, the storyboard and every prompt, and I only talk to you. All drawing goes to Codex through codex exec, with $imagegen in the prompt, one image order, and don't move on until each step passes:
1. Story first. No images until I say yes.
2. Have Codex draw one sample frame. Once I pick the style, start on the real shots.
3. For each shot, have Codex draw the first and last frame. Zoom in on every one, and fix mistakes on the image so they never make it into the video.
4. Before you submit any video, spin up a fresh agent that hasn't seen anything and have it only look for mistakes. Once it says PASS, give me the price.
5. Only run the video after I say yes."
All the drawing runs on Codex, 𝗳𝗿𝗲𝗲 with your ChatGPT subscription, on the strongest image engine out right now, 𝗜𝗺𝗮𝗴𝗲𝘀 𝟮.𝟲. The AI shorts you've seen that don't look AI-made were shot exactly like this.
9
2
51
3,791
Nick CleanCode retweeted
10 GitHub repos that give Claude superpowers (bookmark):
The Claude GitHub ecosystem has gotten out of hand.
Custom Claude skills by Anthropic, full dev workflows, Claude superpowers. All open source, all 100% free.
I went through hundreds. These are the 10 you need.
🧵: Top 10 Claude repos (install today) 👇
2
4
17
5,114
Playing around with tools like this is why I enjoy coding for creative projects. Giving feedback like 'add 4 frames of anticipation' instead of hand-tuning tweens is clever. Might have to give the skills repo a try!
opus 5.5 + remotion + gsap = your own f*cking motion design lab.
give claude code a motion reference and let it break the animation into 6 editable controls:
→ spacing: control how far objects move between poses.
→ typography: adjust text size, spacing, and line breaks.
→ w much blur appears during fast movements.
→ easing: fine-tune how movements speed up and slow down.
→ follow-through: make the tail finish moving after the main object stops.
→ anticipation: add a small backward movement before the main action.
keep everything in one config file.
now you can give feedback like:
“add 4 frames of anticipation.”
“delay the tail by 3 frames.”
“hold the title for 1 second.”
the workflow:
build → render → scrub → adjust → repeat.
the official remotion + gsap integration keeps the animation synced with remotion’s frame clock.
resources:
remotion docs: remotion.dev/docs/gsap/use-…
remotion: github.com/remotion-dev/r…
remotion skills: github.com/remotion-dev/s…
bookmark this and build your own motion design studio, powered by an AI employee.
0
0
0
7