JustHandled Labs
FIELD NOTES / PRACTICAL DELEGATION

What should you actually delegate to a ChatGPT dot?

A clear responsibility brief makes ongoing work easier to review—and easier to keep within the authority you intended.

Conceptual AI-generated illustration: a woman reviews a document while a warm-ivory matte robot waits with the next folder. A metaphor for bounded AI delegation, not Dots controlling physical hardware.
Conceptual AI-generated illustration of delegating routine work while retaining review decisions. The robot is a metaphor, not Dots-controlled hardware.

“Keep an eye on this” sounds like a simple assignment. Until you have to explain what “this” includes, what matters, and whether your assistant is allowed to do anything about it.

That is the practical challenge with an AI that keeps working between conversations. Which job can you stop hovering over? Start with a few release pages, one project or a handful of approved newsletters. Give it a useful result to bring back.

OpenAI describes dots as agents that can keep working between conversations. The continuity is useful when the job is clear. “Help with my business” leaves you supervising a collection of guesses. “Check these three sources and bring me changes that affect this workflow” gives you a result you can actually review. Getting started with a dot explains how to assign ongoing work.

Choose something you can tell is finished

A good first job has named sources and a result you can inspect before a mistake reaches someone else. Let the first complete cycle show you where access or judgment needs work. The following are illustrative assignments, not firsthand Dots tests.

1. Watch releases that matter to your work

Ask for changes on specified public release pages that affect a named workflow. Require original links, dates, what changed and one practical implication. A source it couldn’t open should appear as an access gap. A quiet assistant is reassuring only when the check ran. “Nothing changed” and “I could not open the page” are rather different reasons for hearing nothing.

2. Triage approved newsletters

Limit the job to a sender allowlist in a verified connected mailbox. Ask for a short list of leads, deduplicated by the underlying event, and public corroboration before any factual claim becomes an article. Keep messages, read state and labels unchanged. Exclude personal correspondence and account-specific reports. The output is a research shortlist, not permission to republish private email.

3. Keep project decisions visible

Use an approved plan or planning channel to identify approaching deadlines, missing owners and unanswered questions. Ask for proposed reminders in the review output. Keep sending and changing the shared plan outside this first assignment. The valuable result is a decision you can make quickly, with a link back to its source.

Five parts of a delegation brief: goal, sources, allowed actions, approval boundary and completion evidence. Add timing and stop conditions, then inspect one complete cycle.
A starting brief, not a guarantee. The examples in this guide are illustrative.

Give the job edges

You do not need the perfect prompt. You do need to settle the questions that would otherwise come back to you—or get guessed. Fill in this brief before handing over recurring work.

Goal
What decision or recurring burden should this help with?
Sources and exclusions
Which URLs, folders, senders or channels may it use? Which account? What must it leave out?
Allowed actions
What may it read, compare, draft or update? Name any permitted destination.
Approval boundary
Which actions need your decision first: sending, publishing, purchasing or changing shared records?
Timing and destination
Specify trigger or schedule, time zone, end date and where results belong. Ask it to confirm the saved setup.
Completion evidence
Require sources actually checked, access failures, the output file or link, and receipts for any authorized external action.
Notifications and stopping
Define material changes, when to stay quiet, when to stop and how to handle missing access.

A small first assignment you can adapt

Each weekday at 9 AM America/Chicago, from [start date] through [end date two weeks later], review only [public release URLs]. Identify changes affecting [named workflow], with original links and dates. Save a dated check record in [approved destination], including which sources were checked and any failures. Notify me in [channel] only for a material change or access failure. Do not install software, change accounts, contact anyone or publish. If access fails, report the gap; do not substitute a private source. Confirm the saved schedule, time zone, end date and destination before the first run.

Replace every bracketed field before using it. Define “material” with an example, such as removal of a feature your workflow depends on. Fixed recurring work needs a saved schedule; a conversation about checking something is not proof that monitoring was configured. Tasks and memory describes schedules, supported event monitoring and task context.

Your browser login does not travel with the assignment

A dot’s cloud computer has its own files and browser sessions. Your personal browser login doesn’t sign it in. Work requiring your connected computer needs that computer online with the ChatGPT app open. Supported plugins depend on the connected account, permissions and execution environment; local skills require a connected computer. Computers and apps explains these separate connections.

Before delegating, ask it to confirm the exact source and account it can access. A messaging connection is not inbox access. A listed skill is not proof that its scripts, dependencies or host-specific tools work in your dot’s execution path. Test the actual path before relying on it. Dots are rolling out gradually to eligible accounts; check current access in OpenAI’s Dots documentation rather than assuming the feature is available everywhere.

Look at what arrived, not just the completed badge

Open Activity to inspect the result and errors. Check Scheduled separately for instructions, timing and destination. If an authorized task should change something elsewhere, verify the actual file or external receipt: a completed run alone does not prove successful delivery.

Drafting replies does not authorize sending them. OpenAI’s action review considers instructions, permissions and safeguards; custom rules do not grant missing app access or guarantee error-free behavior. Pausing the main task also does not cancel every delegated task or future schedule. Review and stop those separately. Controls explains the distinctions.

After one complete cycle, ask whether the result helped, what you had to correct and whether the assistant stayed within the job. If you are still hovering, fix the assignment before adding another one.