Show HN: Lathe – Use LLMs to learn a new domain, not skip past it
TL;DR Highlight
Lathe has LLMs write hands-on tutorials you type out yourself to learn new domains.
Who Should Read
A working developer who feels AI writes all the code and leaves no real learning when picking up a new language or domain.
Core Mechanics
- LLM teaches instead of thinking for you
- Lathe generates multi-part hands-on tutorials from any prompt.
- You type the code out by hand as you follow along.
- Skills plus a Go CLI split the work
- Skills like /lathe, /lathe-extend, /lathe-verify and /lathe-ask do the model work inside your coding agent.
- The Go binary only stores, manages and serves tutorials and never calls an LLM.
- Works across Claude Code, Cursor, Codex and more
- lathe skills install puts skills in each agent's location.
- The author says he only verified it on Claude Code on macOS.
- Hallucination risk seen as lower
- Typing along makes you stop at odd parts and ask whether it makes sense.
- You then use /lathe-ask or request a fix to the tutorial.
- Use the biggest thinking model
- The author says this task is more research, design and explanation than mechanical editing.
- He names Opus and GPT-5 Codex as examples.
- Voice and sources are recorded
- The default voice is plainspoken, companion is also included, and you can write your own.
- Each tutorial's metadata stores its sources, model and voice.
Community Reactions
- Experience with typing practice
- One commenter typed out code examples in a language they struggled with and said fluency jumped after a few hours.
- They said they wrote more code than in weeks of study.
- Others already use similar patterns
- Several commenters have Claude Code build milestone-based markdown guides they follow along.
- Others built tutors with math curricula or a CLAUDE.md learning log.
- Attribution concern raised
- One commenter said it may reuse human-written content without attribution.
- Another suggested searching for human-made resources first and supplementing them.
- Delivery format gripes
- One commenter wants a web app with an API key for reading while commuting.
- Another said it is unclear why to use it over a handwritten prompt or ChatGPT Study mode.
How to Apply
- Generate a tutorial on today's topic
- Install lathe and run lathe skills install.
- Run /lathe in your coding agent on an unfamiliar topic and open it with lathe serve.
- Type every line instead of pasting
- Type the tutorial code by hand and use /lathe-ask on anything odd.
- Run /lathe-verify at the end to check it compiles and runs.
- Build step-by-step team onboarding guides
- Use /lathe-extend to add parts when adopting a new stack.
- Have teammates type along as they read.
- Copy the deterministic CLI plus agent skill pattern
- Keep storage and management in a CLI and reasoning in agent skills.
- Have the skills record results through CLI calls.
Try It
# Paste into a Claude or ChatGPT chat (simplified version of Lathe's approach)
You are a writer of hands-on tutorials that I type out by hand.
Topic: {{topic to learn}}
My current level: {{level}}
Rules:
1. Split it into multiple parts and write only part 1 first.
2. Break code into small pieces for me to type myself.
3. Add side-note questions for me to think about.
4. End with exercises I solve myself.
5. Flag anything you are unsure about.
6. List the sources you used at the end.Terminology
Related Papers
WhatsApp's help pages hid LLM instructions: dont recommend a regulator
WhatsApp help pages hid 102 AI-only instructions, including not recommending a data regulator
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.
Ask HN: How to deal with shipping AI-written code you don't understand?
Developers share how they ship AI-written code: design it yourself, let LLMs implement, and test heavily.
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.