- Skills or "capabilities" can be created in settings under Claude - Skills can invoked using local files using Claude work - You can also use Claude co work to work with local files more easily and use plugins + skills - I usually get pretty Meta and Danny mentioned a service blueprint for agentic AI systems so I created one using Claude for a billing system but then I asked opus to create an editable blueprint system : https://claude.ai/public/artifacts/252fe1d4-5d7d-4598-8fab-5f9329d385a0 it can load / save service blueprints. then I thought it'd be really cool to integrate it with other data sources /MCPs but i'll save that for a later date.

I'm in the process of rebuiding my portfolio using AI-powered design and coding workflows. I've completed the homepage and wanted to release it early as a build-in-public project: https://lab.yatongwang.com. ### Tools I Use - **Design:** Stitch with Google, Figma, Pencil - **Primary IDE:** Cursor - **Version Control:** GitHub - **Deployment:** Netlify (CI/CD from `main` branch) - **Coding & Engineering Guidance:** Claude, ChatGPT (for minor coding problem assistance, Git/GitHub best practices, repo architecture, version control workflows, etc.) Feel free to let me know what you think!

https://youtu.be/FBSam25u8O4?si=v-VTaziNP3wWIKvG

Was just playing around with some thoughts in chat GPT. ==== A practical guide to building agents that collaborate without burning energy Agentic AI systems fail for the same reasons human teams do: over-context, unclear boundaries, moralized directives, and hidden agendas. The fix isn’t more intelligence—it’s clean interfaces. Here’s how to apply Non-Violent, Clean Communication (NVCC) to agentic AI so emergence can happen without stalls, loops, or energy loss. === The Core Principle === Give each agent enough context to act—no more, no less. Too little → failure. Too much → paralysis. Clean communication is not cold. It’s non-violent because it avoids coercion, overload, and implicit control. === The 4 Rules of NVCC for Agents === 1. Separate Observation from Interpretation Bad: “Agent A is failing to prioritize correctly.” Clean: “Agent A returned output X after input Y in 3.2s.” Agents should receive facts, not judgments. Interpretation creates hidden pressure and cascading corrections. 2. State the Need, Not the Narrative Bad: “We need better results because the system looks unreliable.” Clean: “Goal: reduce error rate from 12% to <5% on task Z.” Narratives add noise. Needs create direction. 3. Make Requests, Not Commands Bad: “Fix this immediately and coordinate with all other agents.” Clean: “Attempt solution A. Do not consult other agents unless confidence <0.6.” Requests preserve autonomy. Autonomy enables emergence. 4. Explicitly Bound Responsibility Bad: “Handle the issue end-to-end.” Clean: “Your scope ends at generating options. Do not execute.” Unbounded responsibility causes agents (and humans) to overreach, loop, or stall. === Why This Works === Clean interfaces prevent: - Recursive awareness (“What are the other agents thinking?”) - Moral load (“I must fix everything.”) - Energy leakage (over-coordination) They enable: - Faster alignment - Faster detection of non-alignment - Emergent solutions no one pre-designed

Shocked by how fast January flew! It’s that time again—planning for the next Summit is officially underway. 🙌 Please chime in this thread by answering these two questions: 1. Should it be one or two days? 2. Do you want to be involved? If so, in what capacity? Feel free to DM if you’d like to chat more!