Stay the expert. Let AI do the work.
Great AI results come from two things, and neither is the model: the context you feed it, and the judgment to know if what came back is right.
Non-expert work is heading toward zero. Being the expert never mattered more.
Hey @trq212 or @bcherny not sure if this has been reported but seeing this a lot today with `claude agents` when I navigate back and forth into different sessions.
@coreyganim Would definitely watch this. Curious what you think, could this actually work as a side hustle, or does the fulfillment side end up needing full-time attention once you land a few clients?
@mattpocockuk Love this. Most people never get to the step of actually trying to automate their own work. Does /loop-me store what it learns, or start fresh each session?
@garrytan Spent a year building exactly this. Personal brain in Obsidian, company brain as a shared codex my team reads.
What actually unlocked it: making the context machine-readable, not just searchable. There's a big gap between "AI can find my notes" and "AI can think with them."
Point AI at your weaknesses, not your speed.
Before I build anything, I use it as a thinking partner to stress-test the idea. "What am I missing? Where does this break? What am I assuming that's wrong?"
Most people use it to go faster on things they're already confident about. That's the wrong direction.
@coreyganim Solid foundation. The next unlock: stop using a separate chat window and put questions directly inside the notes.
Drop a comment in any doc. AI responds in context. The conversation stays with the content.
Second brain stops being storage. Starts being a thinking partner.
Context is the multiplier.
Before you type anything, load your context. Your patterns, decisions, what was tried last week. Now ask.
Blank chat users are re-explaining themselves every session. Don't let the gap compound.
Judgment isn't a state. It's a transition.
Most people see the AI output and miss the judgment entirely. It doesn't live in the artifact. It lives in the decisions between versions, what to build, what to kill, when the design was wrong.
My little Claude code `/buddy` was nice to have around for about 2 weeks now he's gone. Pretty sure `/buddy` was unshipped lol
I mean I don't blame the team there was very little point
@toddsaunders Obsidian but teach Claude a few skills. Comments for @.claude @.you chats inline `/comments check` finds and responds. Project ~400 line context files in obsidian you load before any related chat `/project search {name}` and `/project update` to capture feedback loop.
@trq212 Have you noticed recently that when compaction runs Claude thinks that skill tool calls were waiting to be run even though they were from further back in the conversation. Then it runs them in sequence often doing nothing because they weren't actually required.
@meetgranola This is so exciting and it's working great in Claude code but your oAuth flow isn't working in Codex, Opencode. Instead of redirecting it opens the App and dead ends.
@GitHubNext anything on the horizon for better PR reviews of markdown files? Our team has moved to tech-designs (PRDs too) in markdown and we would love to be able to comment on rendered markdown. It would be a game changer in our opinion.
Hmm actually Claude was running an integration test that might have caused an issue. I restarted and tried again and it died very quickly, the same way. So might be a totally different issue. Sorry for the ping
@mitchellh I updated ghostty tip today and this afternoon got the 45gb usage crash again. Hasn't happened on "tip" till today. Running multiple Claude codes