Paper, Screenshots, Web clips. All in One Knowledge Graph.

01Why I built it

I kept learning things I never saw again.

Articles, talks, screenshots and paper notebooks ended up in folders I never reopened. I wanted one place that files everything by topic and turns it into notes I can study. My first version quietly stopped for 24 days while reporting success, so I rebuilt it to run on its own, with every result checked.

02What I value

Four rules the system follows.

1. 1

   Keep the original

   Every source is archived word for word. Summaries sit on top and never replace it.

2. 2

   Verify every result

   A script confirms each capture reached the vault and is cited. A finished run proves nothing by itself.

3. 3

   Scope autonomy by risk

   The AI works alone where code can check it, and stops for me where it can't: handwriting.

4. 4

   Least privilege

   The AI reads strangers' web pages, so it gets file tools only. No shell, no web.

03Architecture

Four ways in, one AI router.

Every source writes the same record into one queue. A Claude agent with file tools only archives each capture word for word and files it into topic notes, and a script checks that it landed and is cited. Handwritten pages wait for my review before anything routes.

Claude Code wrote the code; the decisions and reviews were mine.