When your dot needs steering

You do not need a new prompt strategy every time something goes wrong. Name the problem, give the missing detail, and check the next result. These are suggested messages, not transcripts or tested outcomes.

It missed or misunderstood the context

Say: “You treated Friday as agreed. It was only a suggestion. Use Tuesday’s approved note and show what changes.”

Check: Which source and date did it use? Ask it to restate the current goal. Selected conversation context, ChatGPT memory, and dot’s own notes serve different purposes; none guarantees complete recall.

Next: Look for a revised conclusion that reflects the correction. Acknowledging your message is only the beginning; check the affected work too.

It cannot reach the information

Say: “What information or access do you need, and why? Could we begin with an extract I share?”

Check: The connected account, available plugin, permitted actions, and any sign-in request. If the task uses your computer, check that it is connected, online, and has the ChatGPT app open. Messaging your dot does not grant app or computer access.

Next: It may ask for a connection or a step you must complete. Share only the material the job needs; use the product’s private sign-in flow for credentials.

Updates are too frequent or too detailed

Say: “That’s too much detail. Keep routine progress in the checklist. Tell me only about a new blocker or a decision I need to make.”

Check: Whether the unwanted updates come from this responsibility or a separate automation. Writing style belongs in the conversation. A saved automation may need its own notification instructions changed.

Next: Inspect the next update. Tighten the example if it still misses the mark. See preferences versus permissions.

Work is unfinished, or “done” is unclear

Say: “What’s finished and what’s blocked? Show the result so far, checks performed, and anything still waiting on me.”

Check: Open the relevant task in Activity on desktop. Follow its files, results, and requests for input. Compare the result with the agreed goal and original records.

Next: Resolve the specific missing decision or revise the goal. For important actions, ask dot to show the result or confirmation, and check it before relying on it.

A schedule or monitor was discussed but not confirmed

Say: “Show what you actually set up: source, instructions, timing or event trigger, notification destination, and end condition. If nothing is saved, tell me.”

Check: Fixed-time work should appear in Scheduled with the intended time zone and duration. Event monitoring requires support from the connected service and confirmation of the particular event being watched. Connecting a source alone does not establish a monitor.

Next: Resolve any missing setup before relying on it. A contextual suggestion is a different possibility: useful information may prompt a question without a fixed schedule, but that is not a guaranteed alert service. More on ongoing work.

Your priority changed, or you want work to stop

Say: “My priority has changed. Put the launch work on hold. Identify its active delegated tasks, schedules, and monitors; stop or disable the relevant ones and report anything still running.”

Check: Pause stops the main task. Inspect and stop relevant delegated tasks in Activity; disable or delete recurring tasks in Scheduled separately. For an event monitor, ask which subscription or task was disabled and inspect the available control. Do not assume a casual “stop” canceled every kind of future work.

Next: Check the resulting statuses. Stopping does not undo completed actions. Keep unrelated responsibilities running if that is your intention. Official stopping controls.

Keep the useful part

Once you have a better result, name what helped: “Keep distinguishing proposed dates from agreed dates.” Start small again when the source or responsibility changes. For existing Codex work, automations, and execution locations, continue to Working with work already underway.