CB

@BradGroux No way that's got to be one of the coolest things I've ever heard

I was there for the first-ever launch of Starship, and it was one of the most awe-inspiring moments of my life. Highly recommend. x.com/BradGroux/stat…

@alexhillman ah shit - I'm still kinda new to Twitter and not in the habit of checking notifications. Just spotted this now. Read through the pinned summary - super cool design. Excited to dig deeper!

@CobyBergman Spent 2 weeks building this back in early December. youtu.be/Wpz7LNI737Q?si… Has scaled flawlessly about to cross 10k memories.

I wanna hear from folks using this technique 30-60 days of daily usage later. Based on my own experience with memory systems, even with opus, this approach ramps up in usefulness very fast then starts to slowly collapse under its own weight without deliberate scaffolding.

@BradGroux !!! So cool.

I shipped a production bug fix last night while at a Ty Myers concert with my niece and her friends. No laptop. No IDE. Just voice messages via CarPlay from my car, and chat messages from my phone. Here's what happened: My OpenClaw AI agent VERITAS runs 24/7 on my Mac mini. There was a bug issue submitted to the GitHub repo for Veritas Kanban, and I asked it to fix a Docker path resolution bug. I did all of this use voice commands from Apple CarPlay with Microsoft Teams, while driving my truck to pick up my niece and her friends for the concert. • It spawned two sub-agents: → TARS (GPT-5.1) wrote the fix → CASE (Claude Sonnet) reviewed it • The code scored 10/10 on all four quality gates checked by four more agents (Code, Functionality, Performance, Security). • 15 files changed. 7 services refactored. 7 agents used. Tests passing. Docs updated. • I said "ship it" via Teams message from my phone while at the concert between acts. • v2.1.3 of Veritas Kanban went live on GitHub. Issue auto-closed. Total time from "fix this" to production release: 26 minutes. I didn't write a single line of code. This is what building with AI agents actually looks like. Not chatbots. Not autocomplete. A team of AI workers that ship production code while you live your life. We're not close to this future. We're in it.

I'm still learning. Feedback and tips are very welcome

So far, I've included: - A list of my top questions/concerns with building an AI OS - A repo security review skill that's helping me leverage existing open source tools with more confidence - My 'Building with AI' guide from August 2025 - in case it's helpful for folks

Figuring out the best way to build an AI operating system for myself, and created the following repo to document the journey and share helpful tools and resources along the way. github.com/Cob-AI/figurin…

@alexhillman That's super cool. Any chance there's a way I could try this out?

One of the most incredible pieces of emergent behavior in my Claude code memory system is that the same agent that generates memories from my session transcripts also now looks at those memories in a few diff layers of context and suggests feature improvements. Literally every 15 mins it suggests 1-10 ways to make itself more useful. Repeats get promoted to beads. Everything else gets filed for later review. Beads get built. System gets better. This is compounding.

@felixrieseberg So cool. Is Opus 4.5 doing all this?

Claude Code doesn't just resonate with developers anymore. Non-technical people are using it to build things. Technical people are using it for non-technical work. The line is blurring. I'm by far not the first to think about this. Multiple teams at Anthropic have been working on "agentic experiences" for months - Claude not just as a chat partner, but as something that helps you do real work. @bcherny nudged me: can we take what we've built internally and ship an early, scoped-down version in a few days? So we took a small team, set an aggressive deadline ("Monday sound good?"), and got to work. @claudeai wrote Cowork. Us humans meet in-person to discuss foundational architectural and product decisions, but all of us devs manage anywhere between 3 to 8 Claude instances implementing features, fixing bugs, or researching potential solutions. For native code, we use local Git worktrees on our local machines. For smaller or web-code only changes, we just tell Claude to go implement it. When someone reports a bug in Slack, we often just @-mention Claude and tell it to fix it. A human (and another Claude) reviews all code before it's merged, but we're now spending most of our time orchestrating a fleet of Claudes and making decisions than artisanally writing individual lines of code. We're releasing Cowork early. It has rough edges. But figuring out what to build is increasingly the hardest part of software engineering - and we think getting feedback early and hearing what users actually need is how we build something truly good.

Credit to Andrej Karpathy for initially seeding the thought. His content wildly sharp.

When knowledge is ubiquitous, the measure of an effective education shifts from being about the amount of knowledge you can acquire to the value of the things you learn how to do.

Finding Claude Code isn't reliably calling OR following skills. Any tips? Eg: Told it to copy an xlsx & modify formatting -> it re-created it missing sheets + formulas Happened twice: 1st: It didn't call the Doc skill 2nd: Told it to use the skill. It did.. but didn't follow it

Got the screenshotted response from Opus 4.5 in Claude Code. My immediate reaction: “He bleeds. He’s not a god. He’s not invincible.”

... insertions, metadata that might be stale. Be paranoid. If any metric doesn't match, that's a clue — investigate until you find the root cause and fix it. Nothing ships until validation passes." Bonus points for pasting the rest of the thread into your prompt as well.

Sum every summable metric (counts, revenue, totals) in both the original and your output — they must match within $1. Go row-by-row, year-by-year, category-by-category. Think through LLM sharp edges: off-by-one errors, string matching with whitespace, column references after...

Trying to refactor a large, messy financial spreadsheet using Claude Code (with docs skill). A few lessons I've learned the hard way + instructions for how to use them to level up your LLM's output: