Chris Gaafary

@trq212 What about when accepting a plan?

i think we might have undersold 1M context tbh, the performance is so so good, I really just don't clear the context window much these days

@trq212 I like agentic engineering

we need a better word than vibe coding man, Claude can create the most beautiful things

@bcherny @michalbrojak @trq212 What about for implementing plans? Claude gives the option to accept a plan and clear context before starting. With the 1 million context window, do you recommend clearing the context when accepting the plan or keeping the context?

@michalbrojak @trq212 Never

What??!! 1M tokens context window?? Insane. Thank you Claude ❤️ But, when to clear it? @trq212 or @bcherny any tips?

@bcherny a feature that I think would be a really great addition to the mobile Claude code and remote control feature is the ability to see any UI of front end projects. For example if working on a website being able to see the changes to the UI. It could be as simple as Claude taking a screenshot and sending it to the mobile client to be able to review. That approach could work for native development with screenshots of simulators. Basically let the user see what Claude sees via screenshots through the mobile app. Food for thought

@bcherny @trq212 Lately the remote control feature for Claude code has been very brittle. It keeps saying my organization hasn’t provided permissions but I’m just a solo dev and I’ve never blocked remote control. Sometimes reauthenticating helps but I find myself reauthenticating all the time to fix it. Any guidance on what’s going on and how to address it?

@dansemperepico If you run in the sandbox and set reasonable allow permissions ahead of time, then you very rarely get asked to accept a permission. Set hooks for non negotiable asks like any rm or git push command.

You guys all run Claude Code with claude --dangerously-skip-permissions right? Because otherwise how in the world can you sit there accepting every single permission when building something?

Boldin is great for DIY investment planning. But tbh more and more I’m using Claude code / Cowork to make and organize everything locally. Also don’t even need to deal with excel. Can just build custom local interfaces with web or native sdks and store everything in a local SQLite database. Ultimately the tools are less useful than just getting the basic behaviors right. Spend less than you earn. Pay yourself first into your savings and investments. Avoid or slow down lifestyle creep. Avoid divorce. Just like a lot of good health advice. Before worrying about what vitamins/supplements to take or seek out expensive or exotic therapies. Eat well, sleep well, exercise a bit. Avoid drugs, tobacco and alcohol. That’s 90% of it for most people.

question for my doctor friends: what are the best sources/tools you use for managing your finances? things like: loans, savings, investing, insurance, etc. i know white coat investor etc. but outside of the financial planners on linkedin are there better tools you use?

@karpathy Just want to say I love your content and public explorations and experiments! Just wanted to share a technique I’ve been exploring (if you haven’t already). I’ve been implementing a sort of continual learning hack by having an agent (using Claude) evaluate its own knowledge on a particular domain (in my case usually an API) without any tool use. I have it effectively document its own knowledge as thoroughly as it can and then audit those assumptions using tools to do research on the web or test its own assumptions and document all of the errors and lessons learned via research about its own knowledge gaps and then package that into a into an agent skills. I use the skill to inform the model that its knowledge cutoff doesn’t sufficient account for a particular topic and to use the skill to supplement its knowledge. I imagine that automating the updates of skills based on new knowledge cutoffs of new models could act as a sort of blunt hack on implementing a sort of continual knowledge. Remove references as they are no longer needed and update skills as the world changes. I think of it like having a recipe written down to reference until my in grained knowledge no longer requires it. Food for thought.

@bcherny Btw, I shared an idea with you to use claude to assess it's own knowledge gaps and build skills to fill those gaps. Food for thought as an addition to Claude Code (if not already explored). x.com/CGaaf/status/2…

btw

@cwhogg Or just a healthcare series period

Someone please start an SF Healthcare Anthropic Series and please invite me to the first installment.

@trq212 @mattpocockuk Yes and consider calling the AskQuestionTool throughout the process to get more understanding of the users goals and intent with the project. In a fresh project scaffold that may not be obvious from the code alone.

@mattpocockuk yeah i like this

I want to make /init more useful- what do you think it should do to help setup Claude Code in a repo?

Seems that on #ClaudeCode v2.1.70 my user level permissions are not working. I’m getting asked for approval on things that should be already approved at the user level. @trq212

for those who have never worked in healthcare, I made a simulation to get a sense of how difficult it is to do something even very simple like order Tylenol. WARNING: this may be infuriating to some providers

@trq212 Loved this tip. Really helped me get a great spec for an app I built in a weekend. Instead of asking Claude questions and expecting answers, flipping it around is really powerful. It’s nice having a tool that helps you really better understand what you want.

my favorite way to use Claude Code to build large features is spec based start with a minimal spec or prompt and ask Claude to interview you using the AskUserQuestionTool then make a new session to execute the spec

Sharing my journey building an AI enabled trauma triaging tool in my latest blog post. gaafary.com/blog/a-trauma-… Link to demo the tool in the post. Built with #ClaudeCode

100% agree about now as the best time to be a builder! I spend more time building new ideas than I ever did before. I applied for the Healthcare Product lead position. Even if not considered for that position, it would be such a privilege to contribute to building the future of AI augmented healthcare. There are just countless opportunities to implement ideas that were just not possible before! It's been so much fun imagining new ways to do things and increase the value of healthcare rather than just trying to cut labor costs.

There’s never been a better time to be a builder — Opus 4.5 & Claude Code keep surprising me in the quality and completeness of the products they can create. So I’m doing exactly that — putting my product founder hat back on and joining our Labs team to be hands-on at the frontier, building products that channel AI toward solving the world's hardest problems. Excited to pass the baton to Ami Vora as she leads the product team in scaling Claude.

@jarredsumner I just want to say that I really love watching how much you and the rest of the claude code team have been proactively seeking additional feedback and detail whenever there are posts complaining about performance of Claude Code.

So many amazing opportunities with a great native macOS experience. For example wouldn’t it be great to just instruct Claude on UI improvements by using PencilKit to provide visual sketches to Claude to explain what you want with a UI or changes to a UI without manually dealing with screenshots?

@twostraws @bcherny Do it! @bcherny let this man improve your macOS app experience! job-boards.greenhouse.io/anthropic/jobs…

I enjoy working with Claude Code on the terminal, but the Claude desktop app for macOS is an abomination. I'm on the verge of flying to San Francisco and following @bcherny around until he or someone else there lets me burn their macOS app to the ground and rewrite from scratch.