Just showed up in X this morning. https://x.com/stitchbygoogle/status/2046624729403142320 Reading the repo now. https://github.com/google-labs-code/design.md

Most people prompt Claude like they're pulling a lever on a vending machine. Input, output, next. Then they wonder why it stays flat. The shift is simple. Treat Claude like a new employee. A collaborator you invested in. Someone you teach how you work, what you care about, what breaks you. Every session becomes an upgrade. The next session feels like Claude shipped a new version. It didn't. You did. I'm chronically dyslexic. Walls of text are invisible to me. I'll scroll past four critical flags and only catch one. So I said it out loud. I told Claude: "I missed three of the four content moments you gave me. I can't read walls." That was the input. Here's what came out of it, written into memory, applied in every session since: Every response ends with a bold decision marker. DECISION, QUESTION, or NO DECISION NEEDED. Content moments go at the top in a dedicated block. Never buried. Max three lines per paragraph. Bullets aggressive. Bold keywords inside sentences so my eyes have skim anchors. Status and decisions never mixed. Skimmable first, decision clean. I didn't get a better model. I taught the model I had. The mechanism Claude has a memory system. User profile, feedback, project context, references. Every correction, every working pattern, every "do this, not that" can be saved to disk and reloaded next session. The weakness becomes the protocol. The friction becomes the fix. The next session opens with it already loaded. That's why it feels like an update. You gave feedback once. It holds forever. The reframe Stop prompting. Start defining how you work. Tell Claude what frustrates you. Tell it what a good response looks like. Tell it your constraints, your context, your weaknesses. Correct it when it misses. Confirm it when it lands. You're not using a tool. You're onboarding a collaborator. The people who get nothing out of Claude are the people who never told it anything. // A<3

I want to start off that I may not be using the right terminology here, so bear with me... The more I interact with claude and other LLMs and then interacting with the people/users that are pretty far away from anything on this community (I do share the youtube channel all the time), the more I see an opportunity to find a way to use these open source models to create a 'co-work' that can be used inside a closed environment (offline) for two reasons: no real issue with token usage/cost (other than the set up and build and you can swap out models) and then most ppl that I connect with that want to learn how to use AI they are not able to use claude or some other model bc of IT and security reasons. I guess my question is, does this make sense or what am I missing? I am in the process of building this for myself to see whats possible but would be interested to get others input and thoughts, because I am not from the developer world, but i like it here, ppl are nice :)

how do you implement ICM safely in a team! I'd love to get this across my organisation but I'm concerned about permissions and access to information. I'm leading the organisation - what I have access to far supersedes what my team should have access to (client details, pricing, methodologies, etc). And how do I protect all of this leaving with staff if they move jobs etc? I mean the better we build these systems, the more of our secret sauce gets encapsulated, right?

The guy who literally built Claude Code just dropped more real, usable wisdom on vibe-coding than most hour-long tutorials combined. Pure signal. Zero fluff. If you're riding the AI coding wave, this is the one you don't skip. Newbie friendly resource that's still relevant.