Capstone — Automate a Real Task From Your Job

🎯 Goal: Take one real, recurring task from your actual job and automate it end-to-end — safely, on copies — so everything you've learned turns into hours saved every week. ⏱️ Time: ~1–2 hours for the first one. Much less every time after.

Modules 08 and 09 had you practice on throwaway and sample files. This is the graduation: the same moves, but on your work. In Module 12 you made a list of candidate tasks — now we do one for real, together.

Before you start: pick the one task#

Open the backlog you made in Module 12 and pick the single best first candidate. A great first real task is:

• Frequent — you do it weekly, or across many files at once.
• Rule-based — you can explain the steps in a sentence or two.
• Low-stakes — a mistake wouldn't be a disaster (no send-to-client, no system-of-record edits on the first run).
• Made of files you already download by hand — PDFs, spreadsheets, exports.

If you're torn between two, pick the more boring one. Boring is exactly what this is for. Don't pick your scariest, highest-stakes workflow for your first run.

Step 1 — Make a safe playground (this is non-negotiable)#

You're about to point an automation at real work files for the first time. The rule from Module 10 now matters for real:

Work on copies. Never the originals.

mkdir -p ~/work-automation/my-task/inbox
cd ~/work-automation/my-task

Now copy — don't move — a small, representative handful of real files into inbox. Start with 3–5 files, not 300. You want enough to be realistic, few enough to check every result by eye.

Before you touch anything, run this safety check:

• ☐ The files in inbox are copies — the originals live somewhere else, untouched.
• ☐ None of them break your company's data rules (Module 10). If you're unsure whether this data is allowed on your laptop like this, stop and ask before continuing.
• ☐ You could delete this whole folder and lose nothing important.

All three ticked? Good. Now nothing you do next can hurt the real thing.

Step 2 — Run the same rhythm you already know#

This is the exact loop from Module 08, now pointed at your own work. Nothing new to learn — just apply it.

Describe the outcome to Claude Code, and ask it to plan before doing anything:

Here's a recurring task I do by hand: [describe it in plain English — what goes in, what should come out]. The input files are copies in ~/work-automation/my-task/inbox. Before writing anything, explain your plan in plain English, tell me any edge cases you see, and don't change or move any file until I say go.

Read the plan. Answer its questions. This conversation is the work.

Then build and preview — dry-run first:

Good. Build it, but run it in preview mode only: show me exactly what it would do (which files, what changes) without actually changing anything yet.

Verify by eye, then let it run for real, then refine by talking to it — exactly as you practiced. If the output's wrong, say expected vs. actual specifically and loop back.

Step 3 — Make it repeatable (so next week is 2 minutes)#

The whole payoff is doing this once and re-running it forever. Bank the work:

This works. Now make it easy to run again next time: keep a dry-run/preview mode as the default, write a short log of what changed each run so I can undo it, and add a note to CLAUDE.md explaining how to run it. Then give me the single command I type to run it.

Optionally, ask it to save the whole thing as a slash command or a note in CLAUDE.md (Module 11), so future-you just says "do the intake cleanup" and it knows what that means.

Save it with Git so it can't get lost:

git init
git add -A
git commit -m "Automate [my task]"

Step 4 — Prove it saved you time#

Do the honest before/after. Write one line in a running note:

Automated [the weekly intake cleanup] — was [40] min/week by hand, now [~2] min.

That's roughly 30+ hours a year back from one task. This note is also gold when your company asks how you're using AI — you'll have real numbers.

Step 5 — Only now, scale up#

Your first run used 3–5 copied files. Once you trust it:

1. Run it on the full batch of copies — still copies, still preview-first.
2. Spot-check a handful of results by eye. Do they look right at scale?
3. When you're confident, use it as part of your real weekly routine — always copies in, always preview before anything destructive.

Then go back to your Module 12 backlog and do the next one. The second task takes half the time. By the fifth, this is just how you work.

✅ You're done when#

• You automated one real, recurring task from your own job, start to finish.
• You did it entirely on copies, previewed every change, and verified by eye.
• It's saved with Git and re-runnable with a single command (or a slash command).
• You wrote down the time it saves — and you know which task you're automating next.

That's the whole course, proven on your own work. Everything from here is just more of the same, on bigger and better tasks.

Next: What is MCP? — optional background for later.