Implementation
Build week: construct the workflow, run it live three times, measure honestly, and capture the evidence as you go.
This lesson is a working session with a checklist, not a reading. Budget two to three hours across the week (the lesson minutes above count reading time; the work itself spans the week) — most of it doing your actual job through the new workflow.
Build (45–60 min)
- 1Write the one-pager skeleton first: stations, [AI]/[ME] tags, checkpoint positions. Design before prompts.
- 2Draft each AI station's prompt from your Module 2 template. Include the 'list any information you're missing' line.
- 3Dry-run on a past instance where you know the right answer. Fix the biggest gap. Dry-run once more.
- 4Write the risk notes: what never gets pasted (Module 4), what always gets verified (Module 5).
Run live × 3 (across the week)
- 1Run 1: time every station honestly, including your review time. Save inputs and outputs (redacted if needed) into your evidence pack.
- 2Run 2: fix the one thing that annoyed you most in run 1. Note what you changed and why — graders love a visible iteration loop.
- 3Run 3: have the output actually used — sent, filed, presented. The workflow isn't real until its output is.
- 4After each run: one line in a log — date, minutes, what the checkpoint caught (even if 'nothing').
Measure like you'll be audited
Your before/after claim is the headline of your briefing, so earn it: the before number is what the task honestly took (average of your last few manual runs, not your worst day), and the after number includes your review time. 'Ninety minutes down to twenty, of which eight is review' is a bulletproof sentence. 'Basically instant now' is not.
Here is my workflow one-pager and my three run logs: [paste]. Act as a tough but fair capstone grader using this rubric: prompts, workflow design, responsibility, evidence & honesty — 25% each. Give me a grade per category, the single weakest point overall, and the one improvement with the best effort-to-impact ratio before I call it final.
Yes — the model pre-grades your capstone about using the model. It's the self-critique pattern from Module 2, one last time. Fix the weakest point; ignore the flattery.
Two common rescues. Output quality flat despite prompt fixes → your task is more judgment-heavy than it looked; shrink the AI's role to the genuinely mechanical slice and let the human stations grow. Time savings marginal → your review is duplicating the checkpoint; move it upstream (Module 5) instead of reviewing twice.