The Delegation Test

By the end of this install, the weekly task you dread most runs from a single line — and you hold a time receipt that says so in numbers: "Before: 2 hours every Monday. After: 15 minutes of review. Hours back this year: 84."

This is the capstone. There is nothing new to learn. You already own the four patterns — a voice skill (Install 1), a scheduled task (Install 2), triggered workflows (Install 3), and a document skill (Install 4). Today you point them at the one task you'd pay real money never to do again, and you measure what happens.

Prerequisites

• Installs 1–4 built and working — you'll reuse them, not rebuild them.
• Install 0.5 (The Firewall) if you're on Lane A and the task touches work data.
• The component-design worksheet and the time receipt template (downloads below).
• An honest number: how long the task takes you now, per week. Guess if you must — but write it down before you build.

Build steps

1. Pick the task. The right candidate is weekly or more frequent, dreaded, and mostly information-processing — gathering, reformatting, summarizing — rather than judgment. If you're torn, let Claude referee:

   Help me choose one recurring task to delegate to you. Here are the tasks I do
   every week and roughly how long each takes:
   
   - [e.g. Compile the Monday ops summary from four team updates — 90 min]
   - [e.g. Prepare the client pipeline review — 60 min]
   - [e.g. Expense coding and chasing — 45 min]
   
   Score each task 1–5 on: (a) how rule-based it is, (b) how much of it is
   gathering and reformatting information rather than exercising judgment,
   (c) how much I dread it. Recommend the single best candidate, and tell me
   which part of that task must stay with me.
   

2. Write the "before" line now. Open the time receipt and record the current cost — minutes per week, measured or honestly estimated — before you build anything. A receipt written after the fact convinces nobody, including you.

3. Decompose the task. Open the component-design worksheet, then let Claude interview you into filling it:

   Interview me about this task so we can turn it into a component. Ask one
   question at a time, no more than eight questions total, covering: what
   triggers the task, what inputs arrive and from where, the steps I actually
   follow, which steps need my judgment, what the finished output looks like,
   and who receives it.
   
   Then produce a completed component design with this structure:
   Trigger / Inputs / Steps (each marked "process" or "judgment") / Rules /
   Output format.
   

   Answer honestly, especially about the judgment steps — those stay yours. The component does the processing; you keep the calls.

4. Match the design to a pattern, then build. On-demand transformation → a skill. Runs on a clock → a scheduled task. Fired by an event (a meeting ends, a request lands) → a reusable workflow prompt. Then:

   Using the component design above, build the component:
   
   - If it is a skill: write the skill file, in the same format as my document
     and voice skills.
   - If it is a scheduled task: write the exact task instruction I should
     schedule, including timing and delivery format.
   - If it is a workflow: write the single reusable prompt I will paste each
     time, with [PLACEHOLDERS] for the inputs that change.
   
   Reuse my existing components wherever they fit: write in my voice, and use
   my document skill for anything formatted.
   

5. Run it against last week. Feed it last week's real inputs (Lane A: Firewall-scrubbed first) and compare the output to what you actually produced by hand. Fix the gaps the way you did in Install 4 — state the correction, have Claude update the component permanently. Two passes is normal; more than three means a judgment step snuck into the process column, so move it back to you.

6. Fill in the receipt. Run the task for real this week, time the review, and complete the time receipt: before, after, delta, multiplied by 48 working weeks. Keep it where you'll see it — and consider posting it. Your receipt is the most persuasive thing you will ever write about AI, because it's not an opinion.

The Two-Lane note

Lane A (personal Claude): choose a task whose inputs are personal, public, or survive anonymization — and run every input through the Firewall skill before it reaches Claude. Some tasks won't fit: if the job is "reconcile named client financials weekly," raw data on a personal account is a line you don't cross. That's a finding, not a failure — write the task on your Install 9 list as evidence for getting the company lane.

Lane B (company Claude Team/Enterprise): pick any task on the list, and let connectors fetch the inputs themselves — the component reads the source documents instead of you pasting them, which is where the biggest time wins hide. Then put the receipt to work: it is a one-page business case, in hours, from a named manager. Whoever owns the budget reads receipts.

Component shipped

You now have one bespoke component doing your most hated weekly task, built entirely from patterns you already owned — plus a time receipt proving what it's worth. Tomorrow morning, when the task falls due, you trigger it with one line, review for fifteen minutes, and bank the rest.