
What happened
The author details a setup where Claude Code reads an entry document called CLAUDE.md, writes results into role-defined Obsidian notes stored in git on a home Forgejo server, and must pass a lint script plus pre-commit checks on every edit.
Why it matters
Because the notes persist after each conversation ends, the next Claude Code session can resume from the dashboard, decisions, and gotchas rather than starting over, and the same notes are readable by the human on a phone.
Summaries like this, in your inbox every morning.
The workflow is built on a clear division between the AI's working memory and its long-term record. Claude Code keeps a separate 'memory' of about 60 items, but the author restricts it to pointers and cautions so it never conflicts with the notes, which hold the actual facts, history, and decisions.
Much of the system traces back to specific failures. A token printed to the terminal led to the rule about never exposing secrets; a double-imported feed that ballooned a processing queue to 220,000 items led to a check-your-sources step; a near-miss on deleting part of a key led to checking lengths. The mechanical lint layer exists because writing rules into a document did not prove they were followed.
The two-machine setup adds a discipline around git that came from collisions: changelog numbers are assigned right after a pull and pushed immediately, after two machines once picked the same number. The author's summary is that the biggest change was not the work itself but that a record of the work now always exists.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Anthropic used Claude to analyze 25 open-source projects (over 5 million lines) and found 493 issues, includin…

Canadian Imperial Bank of Commerce found that its most powerful models, including Anthropic's Opus, can cost $…

Calif used GPT-5.6 Sol to uncover the critical PHP vulnerability PHP2Shell (CVE-2026-109618/CVE-2026-109619)…

Google opened Gemini 4 Argon to the public on October 13, 2026, after a September 30 announcement limited at f…

Using atopile, KiCad, Freerouting and Claude Code, the developer automated circuit description, placement, rou…

OpenAI published 719 write-ups of math results, but its changelog says only 300 / 719 (about 42%) were verifie…
