A pilot runs the same short list before every flight, however many times they have flown. A project you are about to hand to a coding agent deserves the same.
Here are six checks for a Claude Code project. Each takes a few minutes, and each links to the page of Anthropic's documentation it comes from. We read those pages on November 14, 2026.
The six checks
1. Keep the instructions file short
CLAUDE.md is loaded into every session, so every line costs you every time.
The memory docs give a number: "target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence." A long file is not just expensive. It is followed less well.
Run /init to create one. Then go through it line by line and ask whether removing the line would cause a mistake. If not, delete it.
Where the cut lines go
- The code already says it: delete the line. A written folder map goes stale.
- Only some tasks need it: move it to a skill, or to a file under
.claude/rules/. - It must happen every time: make it a hook. See check 4.
2. Approve the commands you trust, once
Without an allow rule, you approve the same test run fifty times.
Use /permissions to add allow rules for the commands you run all day, such as your test, lint and build commands. The permissions docs describe the rule forms.
Keep the rules narrow. A rule that approves every shell command removes the question you most want to be asked.
3. Put your secrets off limits
By default, nothing stops the agent from reading your .env file.
Add deny rules to .claude/settings.json:
{
"permissions": {
"deny": ["Read(./.env)", "Read(./secrets/**)"]
}
}
Deny rules win. The docs give the order as deny, then ask, then allow, and the first match decides. So a deny rule holds even if a broad allow rule sits beside it.
Look for secrets beyond the top folder. In a project with several apps, each one often has its own .env.
4. Add one hook
A line in CLAUDE.md is advice. A hook is a rule.
A hook is a command that Claude Code runs at a set moment, every time. Formatting after an edit is the classic first one:
{
"hooks": {
"PostToolUse": [
{
"matcher": "Edit|Write",
"hooks": [{ "type": "command", "command": "your formatter here" }]
}
]
}
}
Start with one. Pick the rule you are most tired of repeating.
5. Turn the sandbox on
The sandbox puts operating-system limits around shell commands, and it is off by default.
Type /sandbox. It limits where commands can write and which domains they can reach. It does not limit reads by default, and it covers shell commands only.
That deserves its own post: the Claude Code sandbox is off by default.
6. Know which mode you are in
The permission mode decides how often you are asked, so check it before you rely on it.
The permissions docs list the modes: default, acceptEdits, plan, auto, dontAsk and bypassPermissions. The last one asks nothing at all. Keep it for a container or a virtual machine that holds nothing you care about.
One protection is built in. According to the docs, auto and bypassPermissions have no effect when they are set in a project's own settings files, so a repository you clone cannot switch them on for you.
Then look at what loaded
Open a fresh session and run /context. It shows what is in your context before you have typed anything. If the picture is already crowded, go back to check 1.
About this list
These six are the ones we would not skip. They are our selection from the documentation, not an official Anthropic list, and the documentation can change. We are not affiliated with Anthropic.
We run this list on our own projects with JQ Groundcrew, a free tool that checks all six for you. In a project folder, run npx jq-groundcrew. It runs on your machine and sends nothing anywhere.