Opinion: mastering AI needs clear delegation, visibility, and trust limits
The author argues that effective AI use hinges on three boundaries—delegate/do not delegate, show/do not show, and trust/doubt. Over‑delegating to Claude Code can lead to irreversible actions; a case where a thousand‑hour home‑lab task was handed to Claude Code and finished in six weeks illustrates the risk. To manage delegation, the author proposes four levels ranging from read‑only to…
Key points
- Author splits delegation into read‑only, reversible, irreversible, and money/password levels to prevent accidental AI actions.
- The pubmir CLI creates an anonymized mirror of private repos, stripping IPs and usernames before AI consultation.
- Claude Code’s performance drops after 200 lines of CLAUDE.md, and splitting files does not reduce startup context.
Visibility is addressed with tools like the pubmir CLI, which creates an anonymized mirror of private repositories, stripping IP addresses and usernames before consulting an AI. The piece also notes Claude Code’s hallucination behavior and cites a Zenn test where the model refused to fabricate a fictional specification unless explicitly allowed. Six common failure patterns—scope creep, hallucinated APIs, test tampering, infinite retries, context pollution, and silent failures—are listed with mitigation tips. Claude Code’s performance drops after 200 lines of CLAUDE.md, and splitting files does not reduce the amount of context read at startup.
The author recommends structuring AI interaction through roles (CLAUDE.md, rules, Skill, Hooks) and building failover mechanisms, such as switching to OpenAI Codex when usage limits are hit. Comparing agents, Claude Code suits large‑scale refactoring, Codex or Google Jules handle background jobs, while Cursor and GitHub Copilot excel at inline completion. An example of using Claude Code and Codex as private tutors shows how AI can support learning when the human focuses on answering questions. Together, these practices aim to keep AI a helpful assistant rather than a force that pushes users around.
Model page: Claude Code →
The story so far
7 episodes →- Opinion: mastering AI needs clear delegation, visibility, and trust limitsthis story
3 Boundaries Known Only to Those Who Master AI, Not Those "Pushed Around" by It
note.com · 22 September 2026
Loading the full article…
This text was published by note.com and written by n-chan. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Gemini may delete existing text when asked to append, author reports · 1 src
- Google's Gemini outperforms ChatGPT and Claude as Android assistant · 1 src
- Meta launches Muse AI agent in Mexico, first in Latin America · 11 src
- Meta rolls out real-time voice and expressive avatars for Muse AI in the US · 7 src
- Meta improves Muse and Meta AI with Active Memory and Active Focus · 2 src
Comments
via GitHub Discussions