brb adding a claude code hook that exits with error 2 the moment it outputs "comprehensive"

ok so raw thoughts while exploring OpenAI's new Apps SDK the difficult thing right now is with the existing tooling you have to run two servers: one to serve your assets, one to drive the MCP server logic

launched mcp-lite today: github.com/fiberplane/mcp it's a lighter, minimal library to create your MCP servers in TypeScript that can work on any runtimes and can work with any Standard Schema vendors

if you're publishing an npm library today - what is the standard to inform coding agents about its API (is there)? Something like AGENTS.md but that you can embed in the package.

git remote add @tangled.sh

a dear friend of mine launched an independent label for experimental music today and I helped make a fun website for it first album is very very good, go pre-order: www.glinted.info

trying to use atproto/bsky libraries in Cloudflare runtime is not having a great Sunday hacking time

today's discovery has been being able to easily locally integrate changes from 3-4 different unmerged PRs and one local branch. jj definitely makes for a *really* good review tool

ok, `jj` is very nice

use your ai browsers to buy from artists

"hey i got it working can you give my PR a quick look?

Is there a way to programmatically load some context into Claude Code using CLAUDE.md or hooks? Specifically need it to load the WezTerm's pane id where I have a dev server running so that it can independently look up its output using `wezterm cli get-text --pane-id <pane-id>`

always funny when you ask Claude for an implementation plan and it starts listing: "Week 1-2: .... Week 3-4: ..." oh sweet summer child you're gonna be doing this all today

so DSPy is cool

debuggers are so good

blog.tangled.sh/stacking @oppi.li wrote a great post on stacking (and reviewing!) pull requests using #jj-vcs on tangled!

`git rebase --onto` is great and Idk how I didn't know about it earlier so I wrote about it. laulau.land/note/git-fu/

wrote up a note on a little shell script I use to read generated Claude Code reports: laulau.land/note/markdow...

new ink and switch alert www.inkandswitch.com/ink/notes/sk...

one kind of hilarious thing of trying to combine neverthrow with Cloudflare's Workflow primitives is having to write wrappers that throw..neverthrow-returned errors as Cloudflare Workflows uses caught errors to trigger retries there's probably a better way to do this...

taking the neverthrow library for a spin and all of the .andThen() calls remind me of this scene: www.youtube.com/watch?v=oqwz...

Had a whole lot of fun giving my first conference talk at @squiggleconf.com last year. The crew was very supportive and the vibes were immaculate. If you thought “maybe i should try giving a talk”, then saw a CFP, got excited, and then later thought “ehh idk” then you should think again and apply.

doing some first time (now second time) live streaming tomorrow building a Bluesky MCP on Cloudflare - we may or may not figure out auth too, come hang

some interesting historical context re: stories of "chatGPT-induced psychosis" is that when the radio and telephone were invented, there were similar reactions — some people believed they could tune their radio to the frequencies of heaven or received instructions from the dead via the telephone

if English does become the main language of software expression (supplanting formal languages like code) I'll be very interested in linguistic/dialect analysis on how that changes the language itself

I remember seeing a paper that I believe claimed LLMs being more effective at generating code that uses tools than using the tools directly but I can't find it anymore. What is the latest research/experience on this?

realllly hoping civil defense planners all over Europe are taking scrupulous notes and learnings on human behavior and effects of the power outage in Iberian peninsula. I feel like these kind of occurrences might be more common in the near future whether by force of nature or the enemy

you've heard of PDF in an LLM but have you heard of LLM in a PDF? PDFs continue to fascinate me as a concept, I still don't know if I understand them github.com/evanzhoudev/...

another shoutout to HTML featuring this beautiful poem from Adrienne Rich and its analysis www.nytimes.com/interactive/... love being able to follow the author’s train of thought and see the source text side by side

Anthropic’s latest paper on interpretability of LLMs is worth reading for the presentation alone. More HTML, less PDF in science please transformer-circuits.pub/2025/attribu...

revisiting some of my favorite film shots from last year all shot on Olympus XA - one of those magical small items

what if you could send natural language instructions and tool use as an .mdx file to an LLM

the rush of Claude-generating a python script with comment-embedded dependencies to process unstructured input data, running it with `uv run <script>` and getting a succesful output in a neat output.json file

any MCP clients out there that support the "sampling" feature? I'd be keen to prototype something with it but even Claude Desktop doesn't carry it that yet

what a great website for all of the fans of the Sutro tower: vincentwoo.com/3d/sutro_tow...

Your outie always installs the latest version of drizzle-orm. 👍 Your outie resolves lockfile merge conflicts by hand.

vibes >>> evals

Matthew Rothberg a website that deletes itself after it detects it's been indexed by a search engine github.com/mroth/uninde...

i dreamt i saw a @krazam.bsky.social skit about a prepper who’s really into self-hosting today, how’s your monday going

film still #43 Irving’s dream. Filming pov. #severance

Severance S2 is something else so far. Episode 4 is some of the best TV I’ve seen in a while

I raved about @cloudflare.social Workflows earlier and they're great. I just wish I could pass in a custom `onError` callback for some custom retry logic

wrote an API workflow querying an LLM with a bunch of hand(Claude)-rolled retry logic for error responses (setTimeouts and all). Realized I can just use @cloudflare.social Workflows for durable execution and deleted the hand-rolled retry code. Feels good!

OpenTelemetry is a great tool for debugging. Trying to debug OpenTelemetry code is a nightmare from hell.

The hardest part of working with model context protocol server is counseling Claude through his learned helplessness.

great discussion (from 97!) on interfaces. Fascinating that one of the arguments presented is “direct manipulation = for pro users, agents = for normies”. Lately it seems that it’s the pro users that can really make the most out of “agentic” flows and direct manipulation remains key for most ppl

It’s not a memory leak. We designed our app to keep your memory safe from other programs that want to use it.

has there been any research/benchmarks on whether JSON format is the best structured output format for LLMs? Why not uhhh messagepack or whatever else

me: maybe *this* time I'll maintain a clean commit flow me this time:

a reminder that this is as good time as any to turn off those breaking news notifications - you will need that headspace