Many automation failures happen before the first command runs. The agent has a vague goal, stale assumptions, broad permissions, or no agreed recovery path. A small preflight checklist catches those problems while the work is still cheap to change.

A preflight checklist is a short gate before an AI agent edits files, deploys content, restarts a service, sends a message, or updates a scheduled workflow. It should be boring, fast, and reusable.

What the preflight should confirm

  • Goal: the specific outcome, not just the activity.
  • Scope: the files, services, records, URLs, or documents that may be touched.
  • Inputs: the source material, timestamps, versions, or user-provided instructions being used.
  • Authority: whether the agent is still in read-only mode or is allowed to make state-changing edits.
  • Stop conditions: errors, missing context, privacy risks, or conflicting evidence that require a pause.
  • Verification: the checks that will prove the final claim.
  • Rollback: the smallest safe undo path if the change is wrong.

Keep it smaller than the work

A preflight checklist should not become a second runbook. If it takes longer than the task, operators will skip it. Aim for a one-screen note that answers: what will change, why it is safe to proceed, how success will be checked, and when the agent must stop.

For repeat jobs, store the checklist beside the runbook or scheduled task definition. For one-off delegation, include it in the handoff packet before granting write access.

A compact preflight template

Automation preflight
Goal: <desired outcome>
Scope allowed: <surfaces the agent may touch>
Scope forbidden: <surfaces the agent must not touch>
Inputs checked: <fresh source material and assumptions>
Mode: read-only / draft / write-approved
Stop if: <privacy risk, missing evidence, unexpected diff, failing check>
Verify with: <specific tests, parse checks, or live URL checks>
Rollback: <smallest safe undo or restore step>

Checklist for privacy-safe agent work

  • Generalize examples before publishing or sharing externally.
  • Do not include credentials, account IDs, private paths, private network details, or secret values.
  • Prefer dry-run output before write actions when tooling supports it.
  • Review diffs for accidental private context before deployment.
  • Require live verification for public URLs and local verification for generated metadata.

Copyable prompt for safe delegation

Before taking any state-changing action, write a compact automation preflight. Name the goal, allowed scope, forbidden scope, checked inputs, current mode, stop conditions, verification plan, and rollback path. If any item is missing or privacy-sensitive, stay read-only and report the blocker.

Rule of thumb: preflight checklists are not bureaucracy. They are the moment where an AI agent proves it understands the boundary before it receives permission to cross it.