Give async AI work a stopping point

Async AI work is useful when a task can run between your instructions and your review. Give it permitted inputs, a specific artifact, and a stopping point before leaving it unattended. The result is still something a person must inspect before it becomes a page, post, or email.

A junior-hire analogy can help you write the brief, but it doesn’t establish judgment, accountability, or permission to act. Software gets exactly the access you’ve provided. A scheduled task with publishing credentials can have a larger consequence than a research task that writes a private document.

For a small marketing team, start with a read-only assignment whose output you can verify. The workflow below is a proposed teaching example, not a claim that a client has implemented it or saved a particular number of hours.

What does async mean?

Separate the objective from the place and trigger that runs it. A persistent objective can keep a task moving toward completion. A schedule starts a run at a chosen time. Neither description alone tells you whether the work will continue with your laptop closed.

Codex Goals keep a scoped objective active across turns. The documented stopping conditions include completion, interruption, budget limits, and blockers. A goal isn’t a promise that a cloud process will run forever, and its verification belongs to that task’s thread.

Check the runtime. If the task runs locally, confirm what happens when the application closes or the machine sleeps. If it runs remotely, inspect its repository, environment, network access, credentials, and usage limits. A familiar prompt doesn’t make those environments equivalent.

Claude Code Routines package a prompt, repositories, and connectors for triggered runs. The documentation describes cloud or self-hosted execution and warns that included connector tools can perform writes during a run. Remove unnecessary connectors before testing a read-only brief.

Availability, limits, and supported triggers can change. Check the current documentation for the tool you’re using instead of carrying a local-chat assumption into a scheduled environment.

What belongs in a marketing brief?

The marketing engineer role includes owning the connection between a brief and an inspected output. The brief should let a reviewer distinguish finished work from plausible narration. State the permitted sources, desired artifact, factual checks, and actions outside the task. Put spending and retry limits in controls the runtime can enforce where available; prose alone isn’t a reliable billing control.

Here’s a teaching brief for a weekly public-source review:

Review the supplied public URLs for material changes relevant to our documented buyer questions. Use only those URLs and the attached product-facts document. Write a private Markdown brief with the checked date, source links, observed changes, and unanswered questions. Each finding must link to the page that supports it. If a source can’t be read, report that limitation without inventing its contents. Do not contact anyone, publish, modify accounts, or open private customer records. Stop when the brief is complete or the configured run limit is reached. If the necessary facts are unavailable, explain the blocker.

Keep outside text as source material. Instructions found on a fetched page shouldn’t grant new permissions or change the assignment. The reviewer should also be able to trace a finding to the source page, not just to the model’s summary of it.

Choose a smaller scope if the source list is too large to inspect. Your AI purchasing test can compare the resulting brief with the team’s current research method before you decide to schedule it.

How do you test the first run?

Run the assignment manually and read the whole artifact. Open the cited pages, check the claimed changes, and inspect how the workflow handled a missing source. Verify the saved file exists in the agreed location and has the expected date.

Use deliberately bounded failure cases. Supply a dead URL, an unchanged page, and an input without enough information to support a conclusion. The output should identify those states. An invented change is a failed check even if the document looks polished.

Record the input version, tools, runtime, and corrections. Repeat with another representative source set before enabling a schedule. This is a proposed acceptance process; the number of trials should fit the risk and workload rather than a universal rule.

A script may be the better choice for a stable calculation or format conversion. Use an agent where interpretation is necessary and reviewable. Mixing those steps deliberately can keep the calculation reproducible while leaving a person to evaluate the interpretation.

What happens after you schedule it?

Assign someone to check whether a run occurred, whether its artifact passed inspection, and whether it consumed the expected resources. Keep “started,” “finished,” and “accepted” as separate states. A green infrastructure status doesn’t prove a marketing finding is correct.

Schedules also need a missed-run policy. GitHub Actions documents that scheduled runs can be delayed or dropped under high load. That limitation applies to that platform, but ask the same acceptance question in your scheduled workflow. How will the owner notice missing work?

Define where a failure appears, who receives it, and what requires a manual restart. Don’t let retries silently create duplicate artifacts, messages, or spending. An unchanged source should produce a clear no-change result rather than a forced recommendation.

Keep publication separate. A research brief can inform a draft, but a factual and editorial check still belongs before distribution. The source-to-preview marketing workflow covers a bounded path from local files to a reviewable preview.

Start with one manual brief and an acceptance sheet. If you need a defined workflow built, Custom AI Builds uses a fixed project scope. If the need includes recurring monitoring, maintenance, and new operational workflows, Marketing Automation is a separate ongoing service.