TutorialsLet an Agent Run a Real Job for You
Take a task of your own through the full path — finish it once, wrap it into a Skill, schedule it — and submit all four pieces.
This assignment swaps the class case for your own job: finish it once, wrap it into a Skill, schedule it. Submit four kinds of material, all required: the first result files, the Skill proof, the scheduled task with a test run, and an execution record. It does not need to look polished; the path must genuinely work.
Goals
- Pick a repeated job — one you have done at least three times, roughly the same way — not a copy of the class briefing;
- Let the Agent complete it once, and verify the content yourself;
- Wrap that run into a reusable Skill and schedule it.
Four parts
Part 1 · Finish it once
Follow the Lesson 2 flow with a topic of your own: a dedicated workspace, a requirements.md, and the prompt. Change the topic to something you actually care about — policy updates in your industry, product feedback, competitor releases.
| Item | My answer |
|---|---|
| Task name | |
| Work goal (what the AI should accomplish) | |
| Inputs (which files or information it reads) | |
| Outputs (which files, saved where) | |
| Reuse mode (one-off / Skill / scheduled) |
Self-check: if your "outputs" cannot name a file and a format, the task is not clear yet. Get it to something like "one briefing.docx in the workspace root" before starting.
Part 2 · Verify, including two spot checks
| Check | My record |
|---|---|
| Files actually created and openable | |
| Two sampled items match the original sources | |
| Source links resolve | |
| Format meets the requirement (pages, fields, structure) | |
| Issues found and how they were fixed |
Part 3 · Wrap it into a Skill of your own
Referencing the first result, have it wrap the method into a Skill. Name it yourself so the purpose is obvious: Weekly Report Generator, Competitor Tracker.
| Item | My answer |
|---|---|
| Skill name and purpose | |
| Which inputs the user must supply | |
| Process and standards fixed inside the Skill | |
| How to invoke it after creation |
After wrapping, test it on a different topic. Only a new topic proves the method was stored; testing on the original topic cannot distinguish a real invocation from a copy of the last output. Watch for the skill-loading step during execution.
Part 4 · Schedule it and test-run
| Item | My answer |
|---|---|
| Task name | |
| Run time (for example daily at 9:00) | |
| Skill bound | |
| Save location | |
| Test run result (new files produced?) | |
| Local or cloud, and why |
What to submit
| Category | What |
|---|---|
| First result | Files the Agent produced on its own (briefing, report or other) plus a source list, with a line on how to open them |
| Skill proof | Screenshot of the created Skill, plus a screenshot or recording of one run on a new topic |
| Scheduled task | Screenshot of the configuration (time and bound Skill) plus the test-run output |
| Execution record | The four tables above, filled in, including checks and known issues |
Grading
| Level | Standard |
|---|---|
| Pass | All four materials present; task specific and repeatable; the Agent really produced files; the Skill was created and invoked; the scheduled task configured |
| Good | Above, plus sampled content matches sources; the new-topic test passed; the test run produced real files; failures and fixes recorded |
| Excellent | Above, plus the task comes from real work and is in use; the Skill needs only minimal input to invoke; a clear plan exists for checking automated output |
Job type and output format need not match the class case. Four things are judged: is the job genuinely repetitive, is the output trustworthy, was the Skill actually invoked, did the scheduled task actually produce something.
Pre-submission checklist
- I have really repeated this job; it is not invented for the assignment
- The workspace is dedicated; it cannot touch other folders
- I opened every output file instead of trusting "done"
- I checked at least two items against their sources
- I tested the Skill on a new topic and confirmed it loaded
- I ran the scheduled task once and it produced files
- Materials are desensitized: no client data, credentials or internal files
- I noted which steps still need human confirmation
How to submit
- Upload screenshots of your finished work to the course site;
- Pick the result you are happiest with and share it in the course group so others can see what your automation produced.