
What happened
An operator running Claude Code via claude -p logged roughly 45,000 characters of instruction files loaded at every launch on 2026-11-07 and cut the automated-run CLAUDE.md from about 6,100 to about 3,300 characters by moving history and test results to separate files.
Why it matters
Splitting out material that does not need to be re-read on every launch appears to make each automated session start lighter, and counting characters quickly shows which files carry the most weight.
Summaries like this, every morning.
The official Claude Code documentation says one thing; the author does another in their own setup. On the documentation side, the core advice is to write requests down to the target file or function, keep CLAUDE.md under 200 lines, move multi-step procedures into skills whose bodies load only when used, tag rule files with paths so they load only when matching files are touched, match the model and effort level to the task, disable unused MCP servers, and push long outputs to subagents or output-filtering hooks. As context, the official cost page cites an enterprise average of about $13 per developer per working day (monthly $150–250), with 90% of users under $30 a day, while noting this is an API cost benchmark, not a subscription bill or weekly limit number.
The author's own experiment seems to have come from an unattended claude -p setup that could launch up to 10 times an hour. The measurement on 2026-11-07 showed that the instructions loaded at every launch — CLAUDE.md, shared rules, work instructions, the index — totaled about 45,000 characters. Three changes followed: moving history, known issues and test results out of the unattended folder's CLAUDE.md (about 6,100 to about 3,300 characters), rewriting the index about 25% shorter, and instructing later restarts not to re-read rules and the index. The author frames the common thread simply as sorting by whether something must be read every time.
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.
Amazon.com announced on October 29, 2026 that it has started offering Alexa+, its personal AI assistant, in Ja…

The engineer built OKR等級コンパス with Claude and サスケWorks, then tested it using promptfoo; accuracy rose from 16/1…

Nishi-tō's benchmark scored four approaches on 15 real flaws and 15 safe decoys; the Mantis-style pipeline cau…

Testing three small US shops — Flitch Coffee, Chun Yang Tea and Taquería Chabelita — found new Google reviews…

The developer of HACH, a chat-resident AI agent that works across Slack / Teams / Chatwork / LINE WORKS, publi…

EveryInfra's example live dashboard fetched 9 brands' posts in 39 calls (38 charged, 1 refunded) totaling $0.0…
