# Fathom Mode

- **Event:** [Built with Opus 4.7: a Claude Code hackathon](https://cerebralvalley.ai/e/built-with-4-7-hackathon)
- **When:** Apr 21 at 12:00 PM – Apr 27 at 2:00 AM (EDT)
- **Where:** Online
- **Team:** [ziwei ma](https://cerebralvalley.ai/u/ma-ziwei)
- **GitHub:** https://github.com/ma-ziwei/fathom-mode-for-claude
- **Demo video:** https://www.youtube.com/watch?v=Nxz_zD05Id8
- **Gallery:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery/25

Most AI agents fail not because they reason poorly, but because they reason faithfully from a misread intent. Fathom Mode for Claude targets this failure mode, called Intent Decay: at 85% per-step accuracy, a 10-step agent pipeline delivers only 19.7% end-to-end reliability. It is a planning-session protocol that runs before Claude acts, in both Claude Code and Cowork. Opus 4.7 carries the dialogue: it restates its understanding, surfaces missing details, catches technical misconceptions, and asks one focused question per turn. A deterministic Python core handles the rest, maintaining an Intent Graph, the Fathom Score, Causal Provenance, and a zero-LLM compile step that turns the dialogue into a reviewable plan. The key invariant: Claude may infer supporting structure, but user-stated causal reasoning is kept separate from system-inferred structure before any code is written. After one approval, Claude executes from the compiled plan. The demo shows two real tasks: a daily-work review producing a personalized PDF, and a chaos-theory visualization where Fathom catches the user's Lotka-Volterra misconception in the first turn.

---

Markdown version of https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery/25. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt