Map a workflow before building an agent by writing out every manual step you currently take, identifying the trigger, the inputs each step needs, and the output it produces — then review for steps that could fail or need human judgment.
Why Mapping Comes Before Building
The most common mistake educators make when building their first workflow agent is skipping the mapping step and going straight to building. They open Claude, describe the workflow loosely, and expect a working agent to emerge. What usually emerges instead is an agent that works for the easy cases and breaks for the edge cases — because the edge cases were never defined. Mapping forces you to think through the full workflow before you commit any of it to instructions an agent will follow.
Think of it like planning a road trip. You can get in the car and figure it out as you go — but you’ll waste time at dead ends and miss the turn that would have saved you 40 minutes. Mapping the route first takes 20 minutes. The trip is smoother, faster, and less stressful. Same principle applies to workflow design.
A Four-Column Mapping Exercise
Do this with pen and paper or a simple spreadsheet. Create four columns: Step Number, What Happens, Input Needed, Output Produced. Then walk through your workflow manually, filling in each row as you go.
For example, a video-to-article workflow might look like: Step 1 — fetch transcript (input: YouTube URL, output: raw transcript text). Step 2 — extract key points (input: transcript, output: bullet list of 5 teaching points). Step 3 — write article (input: key points + article format template, output: 500-word draft). Step 4 — review (input: draft, output: approved or revised draft). Step 5 — publish (input: approved draft + category/tag info, output: live WordPress post).
Once you have the map, review it for two things. First, dependencies: does each step have everything it needs from the previous step, or are there gaps where new information appears from nowhere? Second, judgment points: are there any steps where a human decision is required — a quality check, a policy call, an approval — that the agent can’t make reliably on its own? Mark those as checkpoint steps.
Take your completed map to Claude and say: “Here is my workflow map. Build me a skill file that follows these steps exactly, pausing at checkpoint steps for human review.” The resulting agent will be far more reliable than one built from a vague description.
What This Means for Educators
The mapping exercise also reveals whether a workflow is actually ready to automate. If you struggle to articulate the inputs and outputs for each step, the workflow isn’t yet defined clearly enough to hand to an agent. That’s valuable information — it means you need to standardize the workflow first, then automate it.
The Simple Rule
If you cannot map it, you cannot automate it. Spend 20 minutes mapping any workflow before you build an agent for it. That map is the specification the agent builds from — and a good specification always produces a better agent than a rough description.