Most AI disasters are decided before the first prompt is sent: secrets in the context, no branch, no scope, no acceptance checks, the wrong model for the job. This checklist takes five minutes and prevents eighty percent of them. Run it before every bounded task in this class.

The card

Copy this into PRE-FLIGHT.md and fill it per task:

# Pre-flight: <one-line task name> — <date>

- Secrets out: no keys/tokens/customer data in prompt, files, or screenshots. `.env` ignored, key lives in env settings.
- Branch: <branch name> created from clean `main`. Work happens here, never on `main`.
- Scope: may change <folders/files>. Must not change <protected areas>.
- Acceptance: 1) <observable fact> 2) <observable fact> 3) <existing behavior still works>.
- Model: <role-appropriate model> because <risk/cost reason>. Reviewer: <separate model or human>.
- Preview/test path: <command, URL, or test> I will run before calling it done.

Six lines. If any line is blank, the task is not ready — the blank is the risk.

Why each line earns its place

Secrets out comes first because it is irreversible. A key pasted into a chat, screenshot, or committed file cannot be un-sent — only rotated. Check .gitignore, strip customer rows to fixtures, redact screenshots to the last four characters.

Branch is your undo button. Part X teaches the mechanics; here the rule is enough: no task touches the working version directly. A branch isolates the experiment so a bad direction dies alone.

Scope + acceptance turn a wish into a job the model can finish (Lesson 36.2). "Must not change" matters more than "may change" — auth, payments, migrations, env settings, and production data belong on the protected list by default.

Model + preview path connect to Lessons 36.3 and 37.2: the right brain for the stakes, and the exact evidence you will run. Naming the command now stops "looks good" from becoming the test later.

Practical exercise

Fill PRE-FLIGHT.md for your next real task — even a tiny one. Time yourself: five minutes max.

Finish line: six filled lines, branch created, no secrets in the prompt.

Verify: hand the card plus the repo to a fresh session and ask "what would you do first, and what would you refuse to touch?" If it matches your intent, the card is tight enough.

Common failure mode: filling the card after prompting "just to get started." The card *is* the start. Prompt from it, not before it.

Check your understanding

1. Why do secrets come first — what makes that line different from the other five? 2. What does the branch protect you from that careful prompting alone cannot? 3. What belongs on the "must not change" list by default?