New Dev Notes 📬 Most engineers treat #AI output as a single answer: prompt, read, copy, move on. The real skill is closing the loop. Wrote about iterating with the agent, rewinding bad branches, and the two skills I built this week to force it chrisgmyr.dev/newsletter/d...
Every time Claude Code does something unexpected, pause and ask it to reflect on the gap. Have it propose a change to your skills, rules, or config. Push it and move on. We're shipping 1-3 harness improvements a day without stopping feature work. The #AI setup improves as a side effect of using it 🚀
The biggest blocker to #AI adoption isn't the tooling. It's confidence. Watched an engineer hesitate to use AI on a shared repo because they weren't sure the output would hold up in review. The tools are ready. The trust isn't. That gap closes with pairing, not mandates 🧠
Most people stuff everything into CLAUDE.md because they're afraid the model will miss something. It grows until the model ignores half of it. The fix is the opposite: load only what's relevant to the current task - trust that less context with higher signal beats more context with more noise 🧠 #ai
💡 Ran an #AI office hours session this week. The biggest takeaway wasn't a workflow or a tool. It was that nobody is far ahead on this. The field is too new for best practices. If you're experimenting at all, you're in the top quartile. Lower your threshold for sharing what you learn
Model weights don't change between sessions. If it makes the same mistake twice, that's not the model's fault. That's a missing rule. Encode the fix in your repo and it never happens again. Prompt engineering is debugging, not conversation #ai
CLAUDE.md and MEMORY.md have hard limits on what they can hold. Most people treat that as a constraint. I think it's a design signal: if your agent needs more context than fits in a static file, you need conditional loading, not a bigger file. Hooks and rules beat a giant instruction doc every time.
The cost of a cold-start #AI session isn't the first message. It's the 30th. By then you've re-explained your project structure, corrected two assumptions, and lost the thread on what you were actually trying to build. Context loss compounds. The fix isn't a better model, it's a better environment
🚀 New blog post: "My #AI Agent Knows What Project I’m Working On Before I Tell It." I got tired of loading the same context files every session. I built a hook that reads my first message and loads the right project automatically. Bit of bash, JSON config, and zero AI inference 👇🏻
Everyone’s celebrating that #AI writes code faster. The bottleneck was never writing code. It's CI that takes 45 minutes, deployment pipelines with three manual approvals, and product decisions stuck in a backlog nobody owns. Speed up the wrong step and you just get a bigger pile at the next one
New episode of Slightly Caffeinated 🎙️ Walked through the full OpenClaw + Obsidian setup and how the content flywheel actually runs. Also talked Claude Code skill evals, API cost management, and TJ's first two weeks at his new gig
New Dev Notes 📬 I've been building startup hooks, context files, and keyword-matching scripts in Claude Code. I read a "Harness Engineering" article and now have a better name for it. Wrote about what that looks like in practice and why most teams haven't started chrisgmyr.dev/newsletter/d...
New blog post 🚀 I wanted transcripts for podcasts I listen to. So, I built a pipeline over an afternoon: URL in, structured @obsidian.md note out. Whisper handles the audio, Claude handles the structure #ai #pkm chrisgmyr.dev/blog/not-eve...
I started an Obsidian vault for personal notes in early 2024. It now runs my content pipeline on Sunday nights, tracks patterns across my work for performance reviews, and drafts my weekly review. I didn't plan any of that. Each piece just solved the next annoying problem 😅
New blog post 🚀 I set up my content pipeline to run on Sunday nights. Woke up Monday with stats synced across four platforms, a weekly review, and a full content calendar - ready for review and final touches. The system improves its own writing over time: chrisgmyr.dev/blog/i-built...
I started calling tech debt "platform investment table stakes" in planning meetings. Same work, same timeline. People actually engage now instead of immediately pushing back 💡
🔥 I built a Claude Code skill at work that monitors your PR for Copilot comments, auto-patches issues, and resolves them. It notifies you when it's ready for a human review. Already saving time every day
Second week on the new DevX team. Shipping paper cut fixes and building the long-term roadmap. Set up Claude and Copilot in more repos and spent time sanding the rough edges on our internal plugins. Tons of great ideas coming from team interviews 🚀
I wrote context files for my #AI agent expecting it to read them one way. It latched onto the signals file and treated session notes as a task list instead 🤯 New Dev Notes on fixing the AI thread problem: chrisgmyr.dev/newsletter/d...
Love the result! Would you be willing to show what this looks like on your desk?
New blog post 🚀 I wired #OpenClaw into my #Obsidian vault so my #AI agent can read project context, run weekly workflows, and pick up where the last session left off. Full setup and what I learned: chrisgmyr.dev/blog/i-wired...
Everyone measures how fast #AI writes code. Nobody measures how long you spend figuring out what to do with all of it. The real tax isn't prompting. It's routing output from tools that move faster than your decision-making can keep up.
@tjmiller.bsky.social said it on the pod last week and it stuck: the cost of building custom software dropped so much that generic tools are the wrong default now. Just build the thing you actually need 🚀 I've started replacing packages and dependencies - only keeping what I need 🔥
Replaced Claude Projects with a self-hosted #AI agent (OpenClaw) this weekend 🚀 My Obsidian vault is the brain, it holds cross-project context, and I can message it from Discord. Blog post coming this week with the full setup.
Early signs are looking positive. Obsidian vault re-org, OpenClaw on DigitalOcean w/ Discord, and Obsidian headless sync for the whole 🧠!🦞🚀
New Slightly Caffeinated is out! 🚀 TJ’s starting a new adventure Monday, and I’m now full-time in DevX. We also chat about the challenges of managing multi-threaded #AI conversations and the cognitive load of juggling AI sessions.
Running multiple Claude sessions at once. One is on the DevX roadmap, two is researching architecture, three is something I threw in before bed. Fast, but messy. None of them can see the others. I'm the one carrying context. Here’s how I’m solving this
chrisgmyr.dev/newsletter/d...
Spent a few hours on a technical design doc, ran it through Claude Code, and got 5 epics + 40 Jira tickets with full context in minutes. Meanwhile, the team that skipped docs to "move fast" spent the week clarifying things in Slack. Docs save time!
I'm excited to announce that I went from doing DevX ~20% of my time to 100% this week on a new DevX team! I’m working on a formal charter and have been gathering ideas from other teams. It will span across #AI, platform, #API tooling, and more. Super excited to work on this full-time!
