John Boxall

Incredibly proud of the @Mobify team! 🙏 @Mobify signs a definitive agreement to be acquired by @Salesforce: mobify.com/insights/sales…

code mode: let the code do the talking (aka, after w/i/m/p) wherein I ponder the implications of every user having a little coding buddy, and every "app" being directly programmable on demand. sunilpai.dev/posts/after-wi… lmk what you think.

@sintaxi (opencode too!) but I am a poor simple creature of habbits 🙃 next time I run into a wall of 429s, I'll give it another go.

@johnboxall Exists on Pi, and you can customize the footer however you like. pi.dev

Still longing for a omnipresent context window % used in Claude Code!

With Anthropic server tools web search / fetch, I find it necessary to include the # of allowed domains & calls in the prompt to avoid tool call errors. Is it just me running into this?

When using @aisdk ToolLoopAgent, make tools small/incremental. Do not depend on a single "large" tool call for results. Make it possible to use partial work. Use .stream() rather than .generate() even if you don't use the streaming output to prevent the connection from being interrupted.

@randomfrequency look not that critical okay, gotta sleep sometime

@johnboxall "Codex, review Claude's code"

me: "claude, give me critical feedback on this PR" claude: *gives critical feedback* me: *closes Claude, ships change*

@FreedomMobile >1+ hours on hold for a simple change that you chose not to make available online. 🐌🐌🐌

It was probably obvious to everyone but me but in the @aisdk using `jsonSchema` with structured outputs doesn't perform client validation unless you specifically ask for it. This is different than Zod schemas which do perform the validation automatically.

in my culture we don't say "I love you". we say "I left a comment on your PR that's a prompt you can just send straight to Claude Code"

I'm getting tried of using Claude Code to debug my Mac's USB-C connection to my Dell Monitor every other week. I'd be more tired if I had to do it myself. When did updating monitor firmware become a thing?

In the history of the world, has anyone ever been happy with the outcome of pressing this button in Google Slides?

Web search & web fetch (might be) all you need.

Folks building agents, evals can be a nice lab/lagging indicator things are working - what leading metrics are you measuring in the field? output/input tokens, turns, tool calls (and tool failures), duration, parallelism ratio ... am I missing anything else obvious?

Now that the cost of code has dropped drastically I feel much less bad commenting in a PR "did you consider throwing this whole thing away and trying it a totally different way?"

When working on agents I seek to optimize the # of tools calls. I route LLM API req/res through @mitmproxy for visual inspection. To go deep, I dump the flow to disk and get Claude Code to build a visualizer. Never been easier to build tools to build tools 🙃

Mission accomplished!

I also like to live dangerously.

I've got reasonable mileage asking Claude Code to review session files in `~/.claude/projects/.../`, look for recurring tool call mistakes and add correct guidance to CLAUDE.md.

Since the interface for the TS Anthropic SDK changes so frequently, Claude Code isn't great at writing code with it. Prompt it to use the internal/unpublished `claude-developer-platform` skill when working with the Anthropic APIs to improve your mileage.