SAFE / SHIP
LANGUAGE English
Safety check

04 / PILLAR GUIDE

AI OPERATIONS / FIELD GUIDE

How to delegate AI development safely
Keep the boundary visible.

AI agents can move implementation, research, and repairs forward quickly. When the working area and public data are unclear, that speed can also make failures larger.

01 / BOUNDARY

Set the boundary before the prompt

Safe operation starts by separating where an AI may work from where it must not go. That decision comes before prompt quality.

Keep publishing, sending, deletion, billing, and permission changes as human decisions, even when the agent handles the preparation.

02 / PERMISSIONS

Keep permissions small

Do not hand over an entire home directory at the start. Define the project, working directory, and the difference between read-only and write access.

See AI permissions: start small for concrete boundaries.

03 / SECRETS

Keep secrets out of the work path

API keys, private keys, personal data, and customer data should not leak into prompts, logs, screenshots, or commits. Design the workflow so the AI does not need to see them.

See Keep secrets out of the work path for a practical checklist.

04 / REVIEW

Tests and review are different jobs

Tests ask whether expected inputs still work. Review asks what changed, whether data is leaving the project, whether links and metadata are correct, and whether the destination is public.

Use A release review checklist before publishing.

05 / RELEASE

Return control to a human before release

Define stopping conditions first. Stop on an unexpected diff, a failed test, or a request to access data outside the project.

Read Decide the stopping conditions first, then run the one-minute safety check.

SCOPE / SOURCES

What this guide can and cannot prove

This is an operational checklist, not a security audit. Its boundaries follow the permission, working-directory, and human-review guidance in the official Claude Code documentation. Google Search guidance also supports the core principle: use helpful, crawlable content and do not expect special AI-search markup to guarantee visibility.

FAQ

Common checks

Should Claude Code or Codex get full access from the start?

Start with a small project and explicit working area. Keep publishing and deletion behind a human gate.

Should an AI read API keys or .env files?

Keep secrets outside the normal work path and out of prompts, logs, screenshots, and commits.

What should I check before asking AI to write code?

Check the change scope, permissions, secret handling, tests, stopping conditions, and human review before release.

Does a passing test mean the release is safe?

Review the diff, secrets, public destination, external requests, and links before publishing.

Check whether your own workflow is on the safer side of the boundary.

Start the safety check →