GenAI

stop hoarding prompts in Notion. your agent can't read them there, and you re-paste the same three paragraphs every session. Anthropic's agent team said it from the stage: stop building agents, build skills. A skill is a folder your agent opens by itself. What actually moves when you make the switch: - your best prompt becomes a SKILL.md. - your repeated fixes become scripts the agent runs instead of rewriting. - your know-how loads only when the task calls for it. The difference shows up in a month. A prompt library grows until nobody can find anything. A skill library compounds: Anthropic's stated target is Claude on day 30 beating Claude on day 1. New teammate, day one: the agent already knows your stack, your style, your no-go zones. That's the part prompt hoarding never delivered. The full 24/7 stack built on this exact idea (agents writing their own skills, a curator pruning them) is below. Bookmark this. Follow @0xGenAi

@Systeo16 Skill improvements are also mentioned in the video. "A skill library compounds: Anthropic's stated target is Claude on day 30 beating Claude on day 1. "

@0xGenAi Skills are the right shape. The hidden cost: you have to maintain them. I patch mine the session I hit a missing step. A Notion prompt rots in silence. A stale skill rots in public, and your agent has no tact about it.

@sonicdr1p exactly

@0xGenAi This skill approach makes way more sense than a growing pile of prompts in Notion. The agent being able to load what it needs when it needs it is the part that actually compounds.

@kimmonismus Sonnet 5 wasn't underwhelming, its release was undermanaged. Real capability gains that didn't get told well. Same team can ship Opus 5 with a better story.

As much as I'm happy that Fable 5 is staying, I'm equally concerned that Opus 5 will be an underwhelming release, similar to Sonnet 5. What gives me hope, however, is that Mythos preview rumors surfaced back in March, and Fable 5 is known to be Mythos with heavy guardrails. Therefore, I expect the next version, Fable 5.6 or at least 6, to arrive soon.

@limalemonnn just do it, bro

@0xGenAi My procrastination is going to make me poor... I need to stop writing prompts by hand immediately.

@misat0x Save it for your grandkids. And study skills

@0xGenAi my notion prompt graveyard feels attacked

@ZentrixHQ Yes. Actually, the lecture isn't new at all. Yet, most people still don't understand how to work effectively with agents and skills.

@0xGenAi This feels like the next evolution of prompt engineering. Static prompt libraries don’t scale, but modular skills that agents can discover and execute on demand make workflows far more reusable, maintainable, and efficient

@ZentrixHQ Real unlock is one login owning chat, code, deploy. That's the pattern that wins, regardless of who shipped it first. Bookmark and try it. Judge in a week, not on the launch tweet.

OpenAI just quietly merged ChatGPT with Codex - and it changes everything about how you build. You now get 3 modes in one app: Chat, Codex, and a brand new agent called "Work" that can act across your files for hours and ship finished output. Powered by GPT-5.6, which has one wild superpower: UI generation that finally doesn't look like every other AI app. But the real unlock is ChatGPT Sites. No auth. No database setup. No hosting headaches. Just tag "sites" → your app is live on the internet with a shareable URL. One click. Rough idea → production-ready product. Same workflow. Vibe coding just leveled up. Download the ChatGPT desktop app and go build something.

@carbonyxxx Very useful, thank you. It’s wonderful to be able to listen to such intelligent people for free.

Ramamurti Shankar - professor of physics at Yale: "no one understands quantum mechanics." take light, dim it way down, shine it at a double slit. by morning the plate has no fringes - just separate dots. each is one photon carrying the exact same chunk of momentum. so a "wave" is really a stream of particles. now it gets stranger: one particle, two slits - and you still get an interference pattern. a single particle knows how many slits are open. put a lightbulb next to the slits to spy on which one it went through - the pattern disappears. observation changes the outcome not metaphorically: the photon you see with hits the particle. from this comes the uncertainty principle - no mysticism, just the geometry of waves. and the punchline: for the theory to work, the wave function has to be complex. without complex numbers, quantum mechanics stops.

@ZentrixHQ Not a war. It's a chain. Claude writes the brief, GPT compiles it to JSON, Seedance renders. Teams pretending to pick one lose to teams that use both in sequence.

📖 SEEDANCE 2.0. THE JSON PROMPT WAR IS HERE: CLAUDE FABLE 5 VS CHATGPT 5.6 Same prompt. Two models. Completely different creative brains. AI generation is moving beyond simple text prompts. The next battlefield is structured prompting — where every detail matters: character, camera, lighting, motion, style, and consistency. ⚔️ Claude Fable 5 ▪ Strong at cinematic storytelling ▪ Better at understanding long creative briefs ▪ Excellent for world-building, scripts, and narrative structure ▪ Great when you need the AI to think like a director ⚡ ChatGPT 5.6 ▪ Stronger prompt optimization ▪ Better at transforming ideas into precise JSON structures ▪ More consistent with technical constraints ▪ Great for workflows, automation, and repeatable outputs The future creators won’t just ask AI for images or videos. They will build AI production systems where prompts become the new code. The winner won’t be the model with the best demo. It will be the one that creators trust enough to build their entire workflow around 📁Tomorrow I will share another interesting review of this prompt. 🔖Below is an article with a complete guide to using Dreamina.

@misat0x "The output is the controls behind it" flips the whole AI-tools debate. Renders are content. Systems are leverage.

Claude Fable 5 built a procedural city in Blender in under 2 minutes... Then it changed the road layout, block density, and building height live. But... the city is not the output. The output is the controls behind it. A nice AI render gives you something to post. A working Blender system gives you something you can keep building on. → More blocks → Taller buildings → Different street layouts → Your own assets inside the same generator That is the part I keep coming back to. The workflow is no longer: prompt → image → start over. It is: describe → build → inspect → fix what is wrong. Claude can handle the nodes, the Blender interface, and the obvious corrections. You still decide whether the city has taste. But the blank canvas is starting to feel a lot less blank. Full article below👇

@Di_Krass_ "Cost down within weeks" overstates it. Open weights help, but inference cost is hardware bound. Community can quantize and optimize serving. It can't build cheaper H100s.

CHINA released KIMI K3 - an open-source AI model. on quality it's on par with the best American models, and in places better. on per-request cost it isn't cheaper than frontier yet, but since the model is open, developers will bring the cost down within weeks. the main risk isn't the technology, it's the strategy: China already did this with solar panels and electric cars - capture the market cheaply first, then raise prices once everyone depends on it. takeaway: the models themselves are becoming a commodity. that's why Anthropic and OpenAI are moving their business to the application layer (Claude Code, Cowork), while cloud providers (Google, AWS) will earn money running the open models.

@Neuron_404 Loops need a stop condition. "Think in goals" only works if the goal has a checkable definition. Otherwise it's expensive middleware.

WHILE YOU WRITE PROMPTS, HIS LOOPS WORK FOR HIM 24/7. a loop is a tiny scheduled program, but with agent intelligence built in. you give it the outcome, not the steps, and it pulls data from your crm, transcripts and email, decides what matters, and acts within guardrails. > instead of a prompt, you set only the outcome > the loop pulls data from your crm, transcripts and email > it acts within guardrails so it can't go rogue > loops launch other loops - teams of agents while you sleep what used to take dozens of prompts now runs on its own in the background. start thinking in goals, not prompts, and your agents start working by themselves. follow me to learn more about what ai can do.

stop testing Kimi K3 in a chat tab. plug it into an agent with hands. The setup is three moves: hermes model -> Moonshot -> log in with the coding plan. K3 becomes the brain, Hermes stays the hands. What that combo does in this video: /learn digests any guide into a reusable skill. Blender MCP turns a prompt into a 3D product promo. three K3 profiles on a kanban board ship a video: director, judge, builder. His coding side-by-sides match the front-end arena ranking: K3's output came out cleaner than Fable 5's and GPT 5.6's, at a fraction of the price. A cheap open brain plus cheap hands is the whole point. The 16-step guide to that stack ($5 VPS, $7 sub, a five-agent pipeline) is below. Bookmark this. Follow @0xGenAi

@misat0x Thanks

@0xGenAi very useful article,bookmarked thx

@0xVibeDino Thanks, I really tried

@0xGenAi Very good and useful information, saved for myself

@0xKorens Thanks, I really tried

@0xGenAi Nice article bro

@scream_crypto Absolutely. The system and the operator controlling it.

@0xGenAi We’re quickly reaching the point where the model matters less than the system wrapped around it

@UpSideCode effective agent systems

@0xGenAi the kanban with three K3 profiles is lowkey the most underrated part. director/judge/builder is just how good teams actually work