The Five-Step Loop

After you assign, do not walk away. Read the plan before you release, watch the process, check against the brief, and keep a method you can rerun with new source material.

Assigning work is not dropping a sentence in a box. A complete handoff has five steps. Whether the product calls it a task, a thread, or a teammate does not change the five. Skip one, and you keep a lucky success you cannot take with you.

After reading, you should be able to answer:

• Which failure does each step catch?
• Why is “read the plan first” cheaper than “let it start”?
• What do you keep at the end — and what do you not?

Each step catches one failure

1. Write the goal and the deliverable. Fill in the five things. This step catches guessing.
2. Read the plan, then release. Make it restate what it will touch, what it will not, and how done will look. If the plan is wrong, change the plan — not a pile of already-edited files. Codex is described as starting from a goal and then acting. The goal has to be shared before the acting.codex-overview
3. Keep the process visible, and interrupt when needed. You are not a spectator. You watch at stoppable points. Details in watch, stop, take over.
4. Check against the brief, item by item. Check what you wrote. Numbers against detail rows, tests against a command, files against a path. If it fails, stop. “Ship it for now” is not acceptance.
5. Keep the method. Save the skeleton that can run again: the five things, the stops, the check. Next time, swap the source material. Grok Bot treats “show it once, replay it” as a product ability. The judgment is older: a success has to be copyable, not a one-off.xai-bot

The plan is the cheapest intercept

People skip step two because starting looks faster. They find out the direction was wrong after the edits. Rework costs an order of magnitude more than a plan. Plans are cheap. Edited originals, sent mail, and commits on main are not.

Keeping is not saving the chat

What you keep is the skeleton: how you point at the object, how you name the result, where the bound sits, how the check runs. Do not treat the whole transcript as the asset — transcripts rot, skeletons do not. Skills, routines, project files: those are product words. What you want is “next time, swap the source.” How to watch, stop, and take over is the next article.