Most people don't know Claude Code has a hidden ~8,000 character limit on skills.
Past ~15 skills, descriptions get silently truncated. Skills stop triggering. No warning. No error.
The system prompt even tells Claude to never use skills that aren't listed. So truncated skills are invisible AND forbidden.
We built an open source fix.
SkillNote: self-hosted skill registry for Claude Code.
→ Collections scope skills per project (no more truncation)
→ Agents rate skills after use (know what actually works)
→ Edit in browser, every session picks it up in 60s
→ Private registry for skills that can't go on GitHub
One curl command to set up.
https://t.co/J6Zb3LRmez
Show Codex a workflow once. Reuse it as a skill.
Record & Replay lets you show Codex a recurring task, like filing an expense report or submitting a time-off request.
Codex turns that demo into an inspectable, editable skill.
You control when recording starts and stops.
today we're launching @Palmier_io, a video editor Claude can edit.
use AI to edit, organize, and generate footage directly in the timeline.
finally, a video editor built for AI.
open-source. mac native. available now.
We built an AI that can draw on your screen.
It's a true personal tutor.
Using Claude Opus we're able to draw polygons, point with pixel perfect accuracy, and walk users through complex steps directly on their screen.
Here's me learning Pythagorean Theorem + FL Studio.
Demo:
HarnessX: a harness that compiles itself.
every harness improvement so far has come from a human editing code by hand.
Anthropic strips planning steps out of Claude Code when a stronger model ships. Manus rebuilt its agent five times in six months, removing complexity each round.
the craft runs on human judgment about what to change and when. HarnessX is what happens when a system makes those edits itself.
the trick is to treat the harness as a first-class object, the way we already treat model weights.
once it's a typed, editable artifact, it can be optimized from its own execution traces.
the framing they use is an operational mirror. evolving a harness maps cleanly onto reinforcement learning.
the harness is the state. an edit is the action. the trace plus a score is the feedback. a new version is the update.
once you see it that way, the failure modes come for free. reward hacking, catastrophic forgetting, under-exploration.
the same problems that break model training show up when a system edits its own scaffolding.
so edits never ship blind. each round, a loop reads the traces, plans a change, writes the edit, then critiques it.
a gate keeps the new version only if it beats the current one on tasks it hasn't seen.
what makes this safe is the structure underneath. the harness is built from typed components the system can swap without breaking the rest.
that is what compiles really means here. every candidate harness is type-checked before it runs.
here is the result that matters. the weakest model improved the most. the strongest barely moved.
an evolved harness closes the gaps a weak model cannot fix on its own. the weights never changed. the environment around them got smarter.
this is the natural next phase of harness engineering. we moved from weights, to context, to hand-built harnesses.
the harness was the last piece we still tuned by hand.
i wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. the article is below.
paper: HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry: https://t.co/L0GeUKCgef
This is actually nuts. 🤯
An open-source model is now responding faster than most paid APIs. 260 tokens per second. 1 trillion parameters. Free to self-host.
Kimi just shipped HighSpeed mode for their K2.7 Code model. Same intelligence. Same capabilities. 6x faster.
180 tokens per second on coding tasks. 260 on shorter prompts. That's fast enough that the response feels instant.
And this is the same model that uses 30% fewer reasoning tokens than K2.6. Less overthinking. Faster output. Better results.
Open-source under MIT license. No invite needed. Join the beta and you're in.
→ https://t.co/eEtMytSo77
→ https://t.co/Rm2IUjDQHM
→ https://t.co/M5c1yWeU5T
Most people are paying $200/month for models that respond slower than this free one.
How I shipped a viral dev tool in 3 days using AI (without building generic garbage)
Last week, loops (https://t.co/d519SVjRVd) went viral and hit 100% of my Vercel/Neon free limits. Here is the exact framework I use to ship high-fidelity products insanely fast with Cursor
1/ Step 1: Build your own custom primitives (elormui -> https://t.co/yHVMHSUYQq)
If you let AI write your components from scratch, you get generic Tailwind templates
I built my own custom component system and CLI (elormui). My primitives are bulletproof, so when I prompt the agent, it writes UI that actually looks designed, not copy-pasted
2/ Step 2: Orchestration > Typing
Vibe coding isn't about lazy code. It's about letting the AI handle the heavy lifting while you direct
I let Cursor agents scaffold the DB schemas, Prisma DB pushes, and API endpoints. I spend my energy on micro-interactions, high visual fidelity, and polish
3/ Step 3: Real-time production hot-fixes
When loops hit Neon's 402 lock, I didn't panic. I had Cursor analyze the serverless logs, write static fallbacks to bypass the DB block, and map out a Supabase migration while I watched it trend on the web
4/ Step 4: Design-to-code is the killer loop
If you can design and you can code, AI makes you a 10x team. You can bridge complex APIs into magical interfaces because you know exactly how the code should behave under the hood
If you're a creative engineer, stop wasting time writing boilerplate. Build your primitives, trust your eye, and let the agent write the syntax
Hermes Agent now supports asyncronous subagents!
The existing delegate tool, which your agent uses to spawn subagents to fan out and do work, no longer blocks your chat!
To access now, `hermes update`, and enjoy!
This is genuinely insane. This guy burned an entire AI model directly into a chip. No GPU. No CPU. No cloud. Just raw silicon.
It's called GateGPT. He took a small AI model and instead of running it as software, he built it as a physical circuit. Gate by gate. On a single chip.
The result: 56,000 tokens per second. At 80 MHz.
Your phone runs at 3,000 MHz. This chip runs at 80 MHz and is still faster than most AI tools you use daily. Because the AI isn't an app running on a computer. The AI is the computer.
No operating system. No code. No internet connection. Just electricity flowing through a circuit that is the AI.
It generates names on a tiny LCD screen. You turn a physical knob to control it. That's the whole thing.
He started at 2,400 tokens per second. Optimized it 28 times over. Now it
does 56,000. On a chip that fits in your palm.
This is AI stripped down to its most raw form.
This is wild. Actually wild.
Three cheap AI models fused together just outperformed GPT-5.5 and Claude Opus 4.8. At half the price.
OpenRouter just launched something called Fusion.
Instead of sending your prompt to one model, it sends it to multiple models at the same time. A judge reads every response, finds what they agree on, where they contradict, and what each one missed.
Then a synthesizer writes the final answer.
The results on 100 hard research tasks:
→ Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro all budget models beat solo GPT-5.5 and solo Opus 4.8
→ That same budget panel landed within 1% of Claude Fable 5
→ At roughly half the cost
Three quarters of the boost comes from synthesis. One quarter from using models that think differently.
One API call: “openrouter/fusion”. It handles everything.
You might not need the most expensive model anymore. You might just need the right combination of cheap ones.
A 25-year-old housewife in Chennai earns ₹250/hour ($3) just by doing her normal housework.
She wears a phone on her head and records herself making coffee, cutting fruit, folding laundry.
These first-person videos get sent to AI companies training humanoid robots to handle real-world tasks. She shoots 90+ clips a day.
Her quote: "Who else will pay you ₹250/hour ($3) an hour just for doing housework?"
She's part of a growing gig economy in India where thousands are doing the same thing, filming everyday life to train the robots of tomorrow.
Killer playbook for multi-Mac agent setup. Great "how to" guide for Agent Cookie to have your primary Mac and your MacMini agent machine be logged into all the same things with magic cookie sync. https://t.co/IGxJEMgTjZ