# How Nathan Lambert actually uses coding agents

RLHF Book, Interconnects

**If you take one thing:** `/gemini-feedback`, Gemini critiques each diagram. Claude Code draws the textbook figures, then a skill sends each one to Gemini for a second model's review.

- **Agent:** Claude Code (cli) ([source, 2026-08-12](https://www.interconnects.ai/p/i-wrote-an-ai-textbook-how-long-until))
- **Agent:** Codex ([source, 2026-02-09](https://www.interconnects.ai/p/opus-46-vs-codex-53))
- **Model:** Claude Fable 5 ([source, 2026-08-12](https://github.com/natolambert/rlhf-book/commit/1042143366d73ff654727e4d18a1d3a21209c3a3))
- **Model:** GPT 5 Pro ([source, 2026-01-21](https://www.interconnects.ai/p/get-good-at-agents))
- **Review:** Lets agents catch their own mistakes, then checks the results himself ([source, 2026-01-21](https://www.interconnects.ai/p/get-good-at-agents))
- **Harness:** Claude Code instances left running on a DGX Spark while he is out ([source, 2026-01-21](https://www.interconnects.ai/p/get-good-at-agents))
- **Rules file:** CLAUDE.md per repo, AGENTS.md symlinked to it for Codex ([source, 2026-07-30](https://github.com/natolambert/rlhf-book/blob/6063b3de36fff7cb57892e99519dbdd7d72f0225/CLAUDE.md#L31))
- **Skills:** gemini-feedback, pre-submit-pr, push-to-pr, run-rlhf-code-experiment, serve-course-lecture, update-pr-body ([source, 2026-09-02](https://github.com/natolambert/rlhf-book/tree/eecc49e1e1daf1be670d7242eb090240b5bb0f04/.claude/skills))
- **Hooks:** PreToolUse changelog reminder on git commit, git push, gh pr create ([source, 2026-03-25](https://github.com/natolambert/colloquium/blob/2e871435b33bb84b3bd767010f637d2c88511184/.claude/hooks/changelog-reminder.sh#L7-L8))

## In their words

> I would have Claude Code navigate to each comment, print the context before and after, and let me know if it was an easy typo fix or something more nuanced.

> For the last few days, I’ve been using both of the models much more evenly.

> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

> I often have Claude Code pass information back to GPT 5 Pro for a deep search when stuck with a very detailed prompt.

> They make mistakes, they catch most of their own mistakes, and they’re fairly slow too, but they’re capable. I can’t wait to go home and check on what my Claudes were up to.


Archetype (our label): Toolsmith. Newest source: 2026-09-02.
Card: https://www.prehistoriq.com/stacks/natolambert/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/natolambert.json