Lesson 5 of 8 · 6 min read

Build it

Set up your task in Claude or ChatGPT, run it by hand, and review the first drafts.

Build in the order from Lesson 4

You have a spec, a list of sources, and a time for the task to run. This lesson puts them together in a fixed order. Install the skill. Connect the sources one at a time. Run the task by hand at least twice. Add the schedule last.

Each step has a check before you move on. After you install the skill, ask for the task by name and see that the skill runs. After each connection, ask the assistant to list what it can see in that source. After each hand run, compare the draft with what you would have written. Only then do you schedule it.

Show it a finished example

OpenAI published a demonstration of a weekly metrics report built in ChatGPT Work. It followed four steps. Connect the data in Google Drive. Give the assistant the measures that matter, their targets, and a sample of a finished report. Review the first draft. Then schedule the task for Thursdays.

The sample report does a lot of work. A description of the format leaves room for guessing. A real past report shows the length, the order, the tone and the level of detail at once. Attach last month’s best report to your first run, and say that it is the model to follow.

Leave the sample out of the skill file itself. A skill that works from SKILL.md alone is easier to move between assistants. If the sample changes how the draft looks, copy the lesson into the spec as a written rule.

Let the assistant learn your week

Charles Cook at PostHog describes three habits that made his automations useful. First, he asked the assistant to propose tasks from his own week, rather than inventing them himself. Second, he built context files: short documents about his team, its terms and its priorities, which scheduled jobs keep current. Third, he preferred scheduled background jobs to tasks he started by hand in a chat.

A context file helps any report. Write one page that names the people, the projects, the abbreviations and what your readers care about. Point the spec to it. The assistant then writes "the Hartley order" instead of "a large order", because it knows what the Hartley order is.

Run it by hand twice

Run the task by hand on two different days, with real inputs. Read each draft as its reader would. Mark every change you would make before sending it.

Then sort the changes. A change caused by a vague instruction belongs in the spec. A change caused by a missing source means a connection or a process needs fixing. A change of taste, such as a word you prefer, can wait. Fix the spec and the sources, then run the task again.

Do not correct the draft inside the chat and move on. A correction in the chat helps once. A correction in the spec helps every week after.

Five checks for every early draft

Read each early draft against five checks. First, does it cover the right period? Second, can you trace every number to a source? Third, are the sections in the order the spec sets? Fourth, does it flag any source it could not read? Fifth, does it state any reason or cause that no source contains?

The fifth check catches the most serious problem. An assistant asked to explain a late shipment may supply a plausible reason that appears nowhere in your sources. The draft then reads well and states something nobody said. If you find one, add a rule to the spec: report what happened, and give a reason only when a source states it.

Keep these five checks next to you for the first month. After that, most drafts pass them, and checking takes a few minutes. When a draft fails one, note which check it failed in your log for Lesson 7.

Then test the scheduled run itself

A task that works when you start it by hand can still fail on a schedule. Users of both Claude and ChatGPT report this pattern. Permission prompts wait for an answer that never comes. Actions that worked in a live chat get blocked in a scheduled run. One user’s scheduled briefing could not reach the calendar or mail when nobody was at the computer.

So schedule one extra run for a time soon after setup, and check it when it fires. Did it start on time? Did it finish? Did the draft arrive where you expected? Is anything waiting for your approval? Only when a scheduled run passes these checks do you rely on the weekly schedule.

Decide where the draft lands

Choose one place for every draft and keep it there. A dated document in one folder works well, such as "Status report 2026-10-16". An email draft in your drafts folder also works. A message to yourself in a chat app works for short drafts.

The fixed place matters for checking. If drafts land in different places, a missing draft looks the same as a draft you have not found yet. Lesson 6 builds on this.

Worked example · Larkfield Supply

Dana builds the status report

Dana sets up the report on a Wednesday, so the hand runs happen before the first real Friday.

  1. Dana installs the larkfield-status-report skill and asks for it by name. The assistant confirms it is using the skill.
  2. Dana connects the Operations folder, the supplier updates document and the team channel, one at a time. After each one, Dana asks the assistant to list the files or messages it can see from the current week.
  3. Dana runs the task by hand on Wednesday, with last week’s report attached as the model.
  4. The draft is close, but it misses a supplier delay that a colleague mentioned to Dana in a direct message.
    What would you do next?What should Dana change?Think it through, then show the answer

    The process, not the draft. The delay never reached a source the skill reads. Dana asks the team to post supplier delays in the team channel, so the next draft finds them. Fixing the draft in the chat would help this week only.

  5. Dana runs the task again on Thursday. The second draft needs two small edits, both about wording.
  6. Dana schedules the task for Fridays at 1 p.m. and adds one extra run for Thursday at 4 p.m. to test the schedule.That answers the question you guessed at: Dana’s report task works every time Dana runs it by hand. Guess: what should Dana do before relying on the weekly schedule?

The Thursday test run arrives on time in the Reports folder, with no prompts waiting. On Friday the first scheduled draft lands at 1:04 p.m. Dana checks it, edits one sentence, and sends it to the leadership team at 2:30 p.m.

Starter prompts

  • First run, naming the skill

    Use the [your-skill-name] skill to draft this week’s [task name]. Attached is a past example that shows the format I want. Follow its length, order and tone. Use only the sources named in the skill. If any source is empty or you cannot read it, say so at the top of the draft. Save the draft and do not send it.

  • Compare the draft with my version

    Here is your draft:

[paste the draft]

Here is what I would have sent:

[paste your version]

List every difference. For each one, say whether it comes from a vague instruction in the skill, a missing source, or a matter of taste. For the first two kinds, suggest an exact change to the skill or to its sources.

Steps for your assistant

In Claude

Checked on October 15, 2026

  1. Install the skill under Customize, then Skills, and add each source under Customize, then Connectors.
  2. In a new conversation, ask for the task by naming the skill, and attach a past example.
  3. To schedule it, describe the task and the schedule in a conversation, review the name, schedule and instructions Claude proposes, then click Schedule. Scheduled tasks need a paid plan.
  4. Or click Scheduled in the left sidebar, then New task, then Set up manually, and fill in the name, prompt, approval mode and frequency.
  5. Check upcoming and past runs under Scheduled in the left sidebar.

A scheduled task keeps running when your computer is asleep, unless it needs files on your computer.

In ChatGPT

Checked on October 15, 2026

  1. Install the skill in ChatGPT Work with @skill-creator, and connect each source as an app.
  2. In ChatGPT Work, type @, choose the skill, and ask for the task with a past example attached.
  3. Ask ChatGPT to run the task on a schedule, for example every Friday at 1 p.m., and confirm the scheduled task it creates.
  4. Check OpenAI’s scheduled tasks page for where your scheduled tasks are listed and how many your plan allows.

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

  • When the data the report needs is not in any connected source. Fix the source first, or the draft will be wrong every week.
  • When the reader expects your personal judgment in every line, such as a performance summary. Draft those yourself.
Recall from memoryYour task works when you run it by hand. What do you check before you rely on its schedule?Answer it without looking back, then show the answer

Schedule one extra run soon after setup, and check it when it fires. Did it start on time and finish? Did the draft arrive where you expected? Is anything waiting for your approval?

Exercise · your project

Build your task to the draft stage and run it by hand twice.

  1. Install your skill and check that it runs when you name it.
  2. Connect each source, and check what the assistant can see in each.
  3. Run the task by hand on two days, with a past example attached.
  4. Sort your edits and fix the spec or the sources.
  5. Schedule one test run and check it when it fires.

Sources