@dexhorthy @DanielleFong made similar decision earlier too - been sticking to the 40% since I read your article about it awhile back!
English
jp
2.3K posts
Adding opus 1m support meant we revised the context warning in humanlayer from “40%” to “100k tokens”
Matt Pocock@mattpocockuk
Doing some experiments today with Opus 4.6's 1M context window. Trying to push coding sessions deep into what I would consider the 'dumb zone' of SOTA models: >100K tokens. The drop-off in quality is really noticeable. Dumber decisions, worse code, worse instruction-following. Don't treat 1M context window any differently. It's still 100K of smart, and 900K of dumb.
English
This is the best way to use Claude Code...
I've been using Claude with Obsidian since December and my AI workflow has gotten 10x better from this alone. The longer you use this setup, the more powerful it becomes as you expand the memory knowledge graph.
DO THIS TODAY
GREG ISENBERG@gregisenberg
claude code + obsidian in under 1 minute
English
jp 리트윗함
totally agree. well said!
I’ve had most success when I do this and take time up front to drive and set the project up for success before gradually increasing agent autonomy
banteg@banteg
after converting a large portion of the codebase to strict types and fail fast, codex actually starts to pick up what we are doing here. the lesson is you need to drive for some time if you want higher quality result. if you take the hands off the wheel early, the agents various misunderstandings will snowball and you get one big clump of slop.
English
so stoked to try this! observational memory is extremely powerful
already love @mastra and the team keeps shipping even more amazing features!
Tyler Barnes@tylbar
🚨 Announcing a new coding agent that rivals Claude Code but with no compaction needed 🚨 The feeling of using it: run your coding sessions forever, don't worry at all, and get shit done! We're calling it Mastra Code, it's powered by @mastra's new observational memory, and we've been using it internally @mastra to do all our work 1/4 🧵
English
Obsidian has officially won the note wars.
Everyone seems to be scrambling to install complex protocols, paying for plugins or built-in AI subscriptions—or just switching apps once again—as their current app fails to catch up to the latest AI developments.
Meanwhile, over in Obsidian, AI integration looks like this:
Point AI at a folder.
And you're done.
Or, if you're like many reading this, don't integrate AI at all.
Either way, with Obsidian, it's your choice.
Link to the vid below in the comments 👇
English
These are the videos I made with Opus 4.6 using Remotion.
It helped me brainstorm & script them, translate our design system into react components and even generated and inserted the sounds using the elevenlabs APIs.
I never looked or edited at a single line of the code.
English
Opus 4.6 is a special model, it really feels like a true collaborator
you might have got a sneak peek at its work earlier this week- the videos I launched were made completely by Opus 4.6 (see below)
Claude@claudeai
Introducing Claude Opus 4.6. Our smartest model got an upgrade. Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive codebases, and catches its own mistakes. It’s also our first Opus-class model with 1M token context in beta.
English
@AlexTamkin i absolutely love this new feature and see so much potential for self-reflection and also improving how teams/fellow collaborators have our workflows evolve over time
English
Want to use Claude Code better?
Excited to share a little project I’ve been working on!
Check out /insights in the newest version of Claude Code.
Thariq@trq212
We've added a new command to Claude Code called /insights When you run it, Claude Code will read your message history from the past month. It'll summarize your projects, how you use Claude Code, and give suggestions on how to improve your workflow.
English
i saw so much value in this when i first ran it last night that i made a skill to allow for archiving different runs of it so i can keep resurfacing new insights and compare them to previous ones over time
absolutely love this new feature
`npx skills add jonathanprozzi/agent-skills --skill insights-archive`
Thariq@trq212
We've added a new command to Claude Code called /insights When you run it, Claude Code will read your message history from the past month. It'll summarize your projects, how you use Claude Code, and give suggestions on how to improve your workflow.
English
Connect your agents with Obsidian vaults and never let your agents go hungry for the right context again.
On a serious note, in most cases, all you need are markdown files to get your coding agents to do what you want. Skills and Notes (both in md) are all my agents use now.
English
i've worked in the centaur modality for the majority of the last year and i'm feeling this very much right now
finding the need to rapidly adapt and automate what i can while still being in the process where it makes sense
roon@tszzl
it’s just so clear humans are the bottleneck to writing software. number of agents we can manage, information flow, state management. there will just be no centaurs soon as it is not a stable state
English
@nummanali this is great! thanks for pointing this out. i was thinking about this while having xhigh work on a detailed feature. i'm still getting used to some of the codex features
English
In case you forgot, or didn't know - Codex CLI supports Sub Agents
GPT 5.2 Extra High is so thorough that it will take hours on a large task if done on it's own
Direct it to use sub-agents and watch it improve the speed and quality of its output
Huge unlock
English
@wtfsayonara yes! i've been testing out a few things with this and also using Grok for research
it's been super helpful for looking at a whole group of tweets and surfacing commonalities
English
@LLMJunky @Dimillian That’s a cool idea! I’ve been experimenting (with some success) with Codex doing this and generating a report doc first but only a single agent and typically reviewing Opus work. Thanks for sharing this!
English
I have a new strategy for this
I ask it to spin up 2 async subagents to review and check one another's work.
They typically create a report rather than just go off and do the work.
To your point, I will add explicit instructions:
Make sure they're real bugs and not a result of the integration of new features.
English
Codex review is really good but you have to be absolutely careful not to blindly ask it to fix any issues it finds. It (not very often) think you introduce a regression when it’s actually a new intended behavior.
English
@BLUECOW009 Agreed. I’m in favor of folks discovering new tools but this is getting to be draining
English
Every single tweet in my timeline is about clawdbot, the algorithm is pure suffering
English