step 1: buy a discount €9 shitty robot toy. use violence.
@huggingface step 2: figure out the pcb, find out it's a mystery soic, and an h-bridge. the advertisefld "matrix LED display" is 2 segmented color LEDs behind a plastic screen.
@huggingface step 3: ignore everything and rewire the h-bridge to GPIOs on an ft232h, so the phone can later drive the single motor that can either rotate the head left, or move the robot forward in the direction it's heading in. love the mechanism!
@huggingface step 4: write some shitty node/web code for speech to text (local parakeet plus silo vad plus a little magic), text to speech (elevenlabs or local kyutai tts), with as the agent harness, and in the future a local LLM (haiku atm). pi.dev
@huggingface step 5: lunch, knight festival, outdoor play, dinner, tooth brushing, nighty night kisses. clean up boy's laser fortress.
@huggingface step 6: fail at figuring out a new "head" for the robot that can carry the smartphone until steffi casually walks in and solves it in 2 seconds.
@huggingface step 7: draw the rest of the fucking owl, mount phone, connect to ft232h, adjust software until crisp. voila: shitty robot.
@huggingface And if you found this entertaining and have money to spare, consider donating to our charity for 🇺🇦 families in 🇦🇹. zero overhead, every cent goes towards €50 grocery vouchers.
we as software engineers are becoming beholden to a handful of well funded corportations. while they are our "friends" now, that may change due to incentives. i'm very uncomfortable with that.
i believe we need to band together as a community and create a public, free to use repository of real-world (coding) agent sessions/traces. I want small labs, startups, and tinkerers to have access to the same data the big folks currently gobble up from all of us. So we, as a community, can do what e.g. Cursor does below, and take back a little bit of control again.
- one or more platforms to submit traces to for hosting. these platforms must be open to anyone
- support in coding harnesses to let people submit traces easily
- filters/classifiers to remove or warn about any PII or sensitive data in a trace before submission. this is the hardest part
i can't do this as an individual. but we can do this as a community.
i'd be personally more than happy to share all my sessions on OSS this way. all the providers i use for this work already got that data, ToS "we won't use it for training" not withstanding.
My magic little /todo prompt is indispensible for local work now. New/existing project? No problemo. Create docs/todo.md folder, add your first sloppy todo. On first start, Claude will generate a tiny in docs/ it can refer to later. project-description.md
Then Claude asks me which todo from docs/todo.md I want to work on. Like in a text adventure, I specify the number of the todo. Claude removes the sloppy todo from docs/todo.md and puts a structured one in docs/todos/work/${date}-${title-slug}/task.md
It then uses parallel agents to gather info and write a refinded todo description based on its findings. Before (left) and proposed after (right).
I can iterate on the refined description with Claude until I'm happy, then it updates the . task.md
A new entry to my popular series "LLM tools for plebs": claude-trace
- Injects itself into Claude Code
- Logs all traffic
- Reconstructs conversations and shows what's going on behind the scenes (system prompts, all tool inputs/outputs, and more)
Some observations. 🧵
Repo:
I don't consider most of this a trade secret, as the "binary" is just a semi-obfuscated Node app. All prompts are contained as plain string literals. Same for the tool schemas, etc. But what's not in there are the tool inputs and outputs.github.com/badlogic/lemmy…
Let's have a look at the WebSearch and Fetch tools. I asked Claude Code (CC) to research the Anthropic TypeScript SDK.
It uses the WebSearch tool, passing a query to it. It then inspects the results and follows links. Cool, right? Well.