When a prompt is enough (and when it isn't)

Teams default to agents and apps because demos look impressive. Then they inherit memory, permissions, retries, evals, and support load they never budgeted for.

A better default: start at the lightest layer that can finish the job safely. Move up only when the light layer fails for a clear reason.

Ask these in order:

1. 01Is the output advisory for a human who will edit it? -> Prompt.
2. 02Is the path fixed (same steps, same checks) every time? -> Workflow / automation.
3. 03Does the work need to call tools, branch, and recover without a human at each step? -> Agent - with hard budgets and ownership.
4. 04Do people need history, sharing, auth, and a stable URL for a recurring job? -> App.

If two answers fight, prefer the lighter one for v1. Complexity is easy to add and hard to remove.

Each tool on /tools is an app because the job is recurring operator work with shared rules:

• LOCRA - sitemap locs vs live HTTP, same PASS/FAIL every time.
• VEXA - robots traps that block indexing.
• RAUM - Search Console Pages-style reasons, structured for tickets.
• KNOT - canonical conflicts.
• GATE - go / no-go before you request indexing.
• THINN - thin hubs and weak index candidates.
• DRIFT - redirect chains into keep / replace / drop sitemap actions.

A single ChatGPT thread can sketch those checks. It cannot give a team a durable URL, CSV export, email sign-in, and identical verdicts next Tuesday. That is when a prompt stops being enough.

Conversely: writing a one-off brief, renaming a cluster of titles, or drafting a ticket body often stays a prompt. Forcing those into an agent is theatre.

• You built an agent and every run still needs a human to finish - you needed a prompt or a checklist.
• You put out an app nobody bookmarks - the job was not recurring enough.
• You automated a path that changes weekly - you needed judgment, not a brittle workflow.
• You cannot explain ownership, timeouts, or failure copy - you are not ready for an agent in production.

Choose boring on purpose. Upgrade when the work proves it.