claude-code-second-brain

An Obsidian vault template where an AI agent (Claude Code, or any agent that reads AGENTS.md) keeps your notes, decisions and sessions — and gets corrected in writing every time it gets something wrong.

I'm not a developer. For a week I ran my notes, research and a small side project through Claude Code inside one markdown folder. It wrote emails, researched websites, summarized videos, built scripts and ran other agents. It also summarized sources "from memory" so smoothly that four claims in one summary were not in the source at all.

Every mistake like that became a rule in AGENTS.md, with the date and the reason. This repo is that system with my personal content removed.

What's different from the other "Obsidian + Claude" templates

• Rules come from failures, not from a best-practices list. See "Rules learned the hard way" in AGENTS.md.

Add yours the same way: what happened, the date, the rule.

• **confidence: on every note.** verbatim, reconstruction or unverified. We measured it: summaries the model

wrote from memory didn't invent facts, they quietly dropped what didn't fit — on 5 files out of 5 we re-checked. The field makes that drift visible, and /lint lists every reconstruction file.

• **Routing by answers:, not by reading everything.** Each note lists the questions it answers; the index is built

from those lists. Cold start is three reads: AGENTS.md, 00-Index.md, one file.