Lesson 7 of 8 · 6 min read

Measure and maintain

Time the task before and after, watch usage, and keep the setup working when the products change.

Time it again, including review

In Lesson 1 you timed your task by hand. Now time it again for two weeks with the automation running. Count everything: reading the draft, checking numbers, fixing sentences, and sending. The review is part of the task now, and it is the part people forget to count.

Compare the two numbers honestly. The METR study from Lesson 1 found that developers believed AI had sped them up while the clock showed the opposite. Your own sense of the time saved can be wrong in the same way. Only the timed numbers count.

If review takes almost as long as writing did, the automation is not working yet. The log from Lesson 6 shows where the drafts need the most fixing. Most of the time, the fix is a clearer instruction or a missing source in the spec.

Decide what the saved time is for

The UK Department for Business and Trade trial found that staff saved time on written tasks. Its report also found no evidence that the saved time led to higher productivity. Time saved on one task can spread across the week and disappear.

So decide in advance what the saved minutes are for, and write it down. Dana uses the freed Friday hour to call two suppliers each week, a task that had kept slipping. A named use for the time is the difference between a saving and a gap in the calendar.

Measure quality, not only time

Time is one measure. Quality is the other, and it is easy to skip. Count the edits you make to each draft and note their cause. Count the errors you caught that would have reached a reader. A falling edit count means the spec is improving.

Ask the people who read the result, once, after a month. Dana asks the leadership team whether the reports have changed. If readers notice a drop in quality, the time saved is not worth it, and the spec needs work before anything else.

Check usage after the first week

Scheduled tasks use your plan’s allowance, and they can use a lot of it. One ChatGPT user found that a background morning briefing had used a full five-hour allowance before the workday began. A Claude user found that scheduled runs had used up a Max plan and two extra usage purchases, with little visible output.

Check your usage after the first week, and again after the first month. Remove tasks you no longer read. If the product lets you choose a model for a scheduled task, try a smaller one for routine runs, and check that the drafts stay good.

Run frequency drives most of the cost. A weekly report costs little. A briefing that runs every morning, or a monitor that runs every hour, costs many times more. Before you schedule anything hourly, ask whether a daily run would do the same job.

When the product changes

The products change often, and changes break working setups. One ChatGPT Plus user found that personal skills still worked in ChatGPT Work but vanished from regular chat after September 18, 2026. Bundling the skills into a plugin made them available again in both. Another user’s scheduled task depended on agent mode, and the user did not know how to move it when OpenAI switched agent mode off.

Keep a one-line record for each automation: what it does, where it runs, which sources it reads, and the date you last checked it. When a product announces a change, the record tells you which automations to test. After any change, run the task by hand once and check the next scheduled draft.

A record can be this short: "Friday status report. Claude, scheduled Fridays 1 p.m. Reads the shipment log, supplier updates and team channel. Checked October 15." Keep all your records in one note, so one search finds them.

Keep the spec as your master copy. If a product retires a feature, the SKILL.md still describes the task in full, and you can install it elsewhere. That is the main reason this course puts the rules in a file you own rather than in product settings.

Share the spec, not the account

Colleagues who see your drafts may ask for the same automation. Share the SKILL.md file, not your account or your connections. Each person installs the skill, connects their own sources, and runs it by hand before scheduling it.

Sharing the file keeps each person’s data in their own account. It also means each person checks their own drafts. A shared automation that one person checks for a whole team puts the checking back on one desk, and the draft rule stops protecting everyone else.

Retire, promote, and add the next task

Retire automations that no longer earn their place. If you have stopped reading the draft, delete the task. An automation nobody reads still uses your allowance and still reads your data.

Some tasks become so stable that they no longer need an assistant at all. One user had a scheduled Claude task turned into a plain script after a month, to stop spending tokens on a job that never changed. If your draft is identical in form every week and needs no judgment, a plain script or a spreadsheet formula may do it more cheaply.

When the first automation is stable, return to your audit and take the second task. The method is the same: spec, sources, hand runs, schedule, checks. The second one takes about half the effort, because the habits are already in place.

Worked example · Larkfield Supply

Dana reviews the first month

Dana kept a log for four Fridays: minutes spent on the draft, edits made, and anything unusual.

  1. Before automation, the report took Dana 45 minutes. With the draft, Dana spent 25, 20, 15 and 15 minutes on review across the four weeks.That answers the question you guessed at: Writing the report by hand took Dana 45 minutes. Guess: how long does reviewing the AI draft take in Dana’s first week?
  2. Most edits in the first two weeks fixed supplier names. Dana added the full supplier list to the context file, and the edits stopped.
  3. One Friday, the draft arrived late because a permission prompt was waiting. The 1:30 p.m. reminder caught it.
  4. Usage stayed well within Dana’s plan, because the task runs once a week.
    What would you do next?Review now takes 15 minutes against 45 for writing. Should Dana keep the automation?Think it through, then show the answer

    Yes. The saving is 30 minutes a week, and the log shows the drafts improving. Dana also checks that the saved time has a use, the supplier calls, so the saving does not vanish into the week.

Dana writes a one-line record for the report, marks the meeting follow-ups as the next task, and starts the spec for it using Lesson 3.

Starter prompts

  • Find patterns in my log

    Here is my log from the last two weeks of an automated task:

[paste your log: date, minutes spent reviewing, edits made, anything unusual]

List where the drafts needed the most edits, and group the edits by cause. For each group, suggest one exact change to the skill or its sources. Then say whether review time is falling, flat or rising.

  • Check my setup after a product change

    This product change was just announced: [paste or describe the change]. Here is my record of automations: [paste your records]. List which automations it could affect and what I should test for each, starting with the one most likely to break.

Steps for your assistant

In Claude

Checked on October 15, 2026

  1. Check how much of your plan’s usage the scheduled tasks consume, after the first week and after the first month.
  2. Open Scheduled in the left sidebar to pause, edit or delete a task.
  3. When you set up a task manually, you can choose a model for it. Try a smaller model for routine runs.

In ChatGPT

Checked on October 15, 2026

  1. Watch your usage limits after the first week of scheduled runs.
  2. Pause or delete scheduled tasks you no longer read, from your list of scheduled tasks.
  3. If a personal skill disappears from regular chat, check whether it still works in ChatGPT Work. Packaging it in a plugin is a reported fix.

OpenAI’s help pages could not be read directly when this was checked. Confirm each step on OpenAI’s current pages.

When not to use AI

  • Keep the task manual if review takes as long as writing it did, after you have fixed the spec.
  • Stop an automation whose drafts you no longer read.
  • Do not hand the checking to a colleague without agreement. The person who checks the draft owns what it says.
Recall from memoryWhen you time the task again with the automation running, what do you count, and why trust the clock over your own sense?Answer it without looking back, then show the answer

Count everything: reading the draft, checking numbers, fixing sentences and sending. The METR developers believed AI had sped them up while the clock showed the opposite, so only the timed numbers count.

Exercise · your project

Keep a two-week log and decide what happens next.

  1. Time each review, including fixes and sending.
  2. Record every edit and its cause, in a single running note.
  3. Check your usage at the end of the first week.
  4. Write the one-line record for the automation.
  5. Decide to keep, fix or retire it, and pick your next task from the audit.

Sources