Who it is for
Independent builders using AI agents for development and Web production.
AI DEVELOPMENT / FIELD GUIDE 01
Let AI build.
Keep control.
The more you delegate to Claude Code, Codex, or another AI agent, the more important it becomes to define where it may work and where it must stop.
AI OPERATIONS SAFETY CHECK / 1 MIN
Before you hand over the terminal, check permissions, secrets, review, testing, and logs with five questions.
The games on this site were also built under these boundaries.
01 LOCAL FIRST02 HUMAN GATE03 PROOF LOG
DEMO RUN / SAFE-SHIP-001SAMPLE
CURRENT JOBCheck the safety boundary
BOUNDARY SCAN DEMO / NOT A REAL AUDIT
SCOPE触っていい場所を確認25%
01RULESWrite the constitutionDONE
02GUARDLock destructive actionsDONE
03LOGRecord the evidenceRUN
04HUMAN GATEYou decide when to publishWAIT
RELEASE POLICYHUMAN GATE required
5questions
one minute
one minute
0answers
sent to a server
sent to a server
1 gatehuman approval
before release
before release
This is a self-check from real AI-assisted development logs, not a security audit.
AI-READABLE / START HERE
Put the answer
where it can be checked.
SAFE / SHIP is a safety check and field guide for people delegating Web development to Claude Code, Codex, and other AI agents.
Read the AI visibility guide ↗What it does
A five-question, ten-point self-check finds the first boundary worth fixing.
What it does not promise
It does not inspect your files. A full score is not a safety guarantee.
01 / QUICK DIAGNOSTIC
How far can your
AI development setup go?
This is not a test to pass. Answer five questions and find the first boundary worth fixing.
00 / 05
01
You limit where the AI can work
The working directory and off-limits locations are clearly separated.
02
Secrets stay away from the AI's reading path
API keys, .env files, private keys, and personal data are not fed into prompts.
03
Tests and builds run after every meaningful change
You keep evidence that the change works instead of trusting a feeling.
04
A human reviews and approves before production
The AI can create a change without being allowed to publish it alone.
05
You keep a record of what changed and why
A reproducible log keeps the next failure from starting at zero.
FREE SAMPLE / START HERE
Three rules are enough
to start safely.
Use deny, ask, and allow as a small operating constitution:
- deny = destructive actions and secret-file reads
- ask = actions that change intent, publish, send, or spend
- allow = reading, inspection, status checks, and tests
Start with the minimum boundary today. Grow it from real failures.
deny / ask / allow02 / THE WORKFLOW
Before adding more AI workers,
design the handoff.
A reliable line decides what to pass forward, where to stop, and what evidence to leave behind.
01
RULES
Write the constitution
Define the places and actions that need a human check.02
GUARD
Lock the dangerous paths
Use deny and ask to make the boundary mechanical.03
LOG
Leave evidence
Keep the test result and the change record.04
HUMAN GATE
Decide the release
The publish button stays with a person.03 / FIELD NOTES
Failures leave clues.
Keep the useful ones.
05 / ELSEWHERE
Things built,
and the person building them.
Play browser games and follow the development record behind this field guide.
READY WHEN YOU ARE
Give your setup
one careful minute.
04 / FAQ
FAQ
Does this check guarantee that my setup is safe?
No. It is a self-check that helps you find the first gap. It does not inspect your files or environment.
Are my answers sent anywhere?
No. Answers stay in this browser and are not sent to a server. The site analytics measures page visits, not your answers.
Who is Nanashi?
A student and independent builder documenting AI-assisted development, failures, and the boundaries that made delegation workable.