Hacking with Claude on a $27 smart watch
TL;DR Highlight
Built a custom watch face for a $27 PineTime in hours using coding agents
Who Should Read
A backend or app developer who wants to try coding agents on embedded or hobby hardware
Core Mechanics
- Open firmware and a simulator suit agents
- The PineTime costs $27, runs open source InfiniTime firmware, and is well documented.
- The InfiniSim simulator gives the agent a fast feedback loop on the computer.
- Most work was done in OpenCode with open-weight models
- The post keeps saying "Claude" but most work happened in OpenCode.
- Models used were Kimi K3, K2.6, and DeepSeek v4 Pro and Flash.
- First attempt from a photo was rough
- The agent was given a Casio watch face photo and an existing InfiniTime watch face as a base.
- It guessed text sizes and positions, so elements overlapped and were unreadable.
- Fixed one element at a time
- To save tokens, the author gave isolated tasks and concrete next steps.
- Each text element was finished before moving to the next.
- Static parts became a fullscreen background image
Community Reactions
- Opinion that PineTime's RAM is a painful bottleneck
- A reply says the upcoming "pro" version may fix this and links a Pine64 post.
- Experiences hacking other devices with AI
- One commenter built a random I-ching "kookwatch" on a lilygo t-watch with Codex.
- Others built a Garmin calorie tracker with Claude and revived a 10-year-old Pebble watch face.
- Complaint that the "Claude" naming is confusing
- A commenter found it misleading that the post says Claude throughout despite using other models.
- Another asked whether Claude is becoming the Kleenex of agents.
- Info that PineTime has no microphone
- So voice control is impossible.
- Several commenters asked for other hackable watches, and one suggested the Waveshare ESP32-S3 AMOLED 2.09” watch.
How to Apply
- Have your agent get InfiniSim building today
- Clone the InfiniSim repo and ask the agent to get a build working.
- Once the simulator runs, you can experiment without owning a PineTime.
- Give the agent a screenshot feedback loop
- For UI work, let the agent see simulator screenshots directly.
- This cuts manual description and token use.
- Split big tasks into one element at a time
- Give concrete next steps for a single element.
- Move on only after it is right.
- Have the agent write lessons into AGENTS.md
- At the end of a session, ask it to summarize problems and fixes into AGENTS.md.
- The next session or another person avoids the same pitfalls.
Try It
# Paste into Claude Code or OpenCode as-is (the article's method turned into a prompt)
Clone the InfiniSim repo (with the InfiniTime submodule) into {{work folder}} and get it building in this environment.
Once it builds, use an existing watch face as a base to make a watch face similar to {{reference image}}.
Feel free to use sub-agents for well-scoped tasks.
Fix only one text element at a time, and move on to the next after I confirm.
At the end, summarize what you learned and what blocked you in AGENTS.md.Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.