Athos

Athos turns how you work into plain files your AI agents follow. Teach a task once in any agent (Claude Code, Claude Desktop, Codex, Meta's Muse Code, OpenCode) and it becomes a page of your own steps, criteria and exceptions, with branches that open only when the case needs them, kept in ~/.athos on your computer. Next time you ask as you always do: the agent follows your way of working, you watch each step tick in a panel that shows both the strategy (each branch with its condition, taken or not needed) and the content (the step being applied, with when it applies and why), and you can change a step on the spot. The problem: every new chat forgets your checklist, so professionals (lawyers, analysts, journalists) re-explain it or get generic answers. Measured with athos bench: Muse Spark with the person's know-how got 25 of 25 runs right vs 2 of 25 without, on five professional tasks; Meta's Muse Code and OpenCode followed 10 of 10 stages with it (6 of 10 without). A public gallery of ways of working is measured for every profession in it (an accountant, a lawyer and a support team on Muse Spark; a newsroom and a lab on Muse Spark Open: 5/5 with the page, 0/5 without), and anyone can share theirs by asking their agent. Open source, Apache 2.0. Pre-existing: a one-day prototype (Oct 25), rewritten during the event; the submitted repo holds only work done during the hackathon.