Difficulty: Advanced · Estimated time: 20 min
What You'll Learn
How to turn a recurring piece of work into a documented workflow with a trigger, defined inputs, numbered steps, an output and at least one human checkpointThe difference between a good prompt, a good agent, and an actual workflowA five-part template you'll reuse through the rest of this levelWhy This Matters
Guide 05 introduced the basic idea of breaking work into steps. By now you've written solid prompts, managed context well, and built at least one agent. The gap between "a useful agent" and "a workflow" is structure: a workflow is written down clearly enough that a colleague, a client, or future-you could pick it up and run it correctly without you in the room. That's what makes it reusable rather than a one-off trick you only remember how to run yourself.
The Concept
A workflow has five fixed parts:
Trigger — what starts it (a new brief lands, a weekly date, a client sends assets)Inputs — the exact material needed before step one can beginSteps — numbered actions, each with its own input and outputOutput — the final deliverable, in a specific formatCheckpoint(s) — where a human reviews before the work goes out or continuesThe key habit is treating each step as a self-contained unit with its own input and output, rather than one long prompt trying to do everything. That's what lets you swap, fix or improve a single step later without rebuilding the whole thing.
Step-by-Step
Name the trigger. Be specific — "client sends a raw call recording" is a trigger; "when I have time" is not.List exact inputs. Name the real material: a transcript, a brand brief, a spreadsheet of names, specific fields — not "the relevant info."Write numbered steps. For each one, state the action, what it needs, and what it produces. Three to six steps is normal; more than that usually means you've merged two workflows.Define the output precisely. Format, length, where it's delivered — "a draft" is not precise, "a 150-word LinkedIn post in a Google Doc" is.Mark the checkpoint(s). Decide which step(s) a human must look at before the work continues or ships. Guide 24 goes deeper on this.Write it down somewhere persistent — a Notion page, a shared doc, a Project's instructions — so it survives beyond your memory of building it.Example
A freelancer runs a weekly "client update" workflow:
Trigger: every Friday at 3pmInputs: this week's completed tasks list, next week's plan, any blockers (from a shared doc)Steps: (1) Claude drafts a client-facing summary from the raw task list → (2) Claude flags anything that reads as a risk or delay → (3) human reviews and edits tone → (4) final version is sentOutput: a 200–300 word email, ready to sendCheckpoint: step 3 — nothing goes to a client unreadNothing here is exotic. What makes it a workflow is that it's written down and repeatable, not reconstructed from memory each Friday.
Try This
Pick one task you already do at least monthly. Write out its trigger, inputs, steps, output and checkpoint using the template below — even if you'll keep running it by hand for now.
Copy This
WORKFLOW: [name]
TRIGGER: [what starts this]
INPUTS:
- [input 1]
- [input 2]
STEPS:
1. [action] — needs: [input] — produces: [output]
2. [action] — needs: [input] — produces: [output]
3. [action] — needs: [input] — produces: [output]
OUTPUT: [exact format and destination]
CHECKPOINT: [which step a human reviews, and what they're checking for]
Common Mistakes
Writing one mega-prompt and calling it a "workflow" — if it isn't broken into steps with their own inputs/outputs, it's still just a promptVague inputs ("the context") that mean you re-figure-out what's needed every timeNo checkpoint at all, so a bad first step quietly reaches the clientOver-engineering a workflow for a task you do twice a year — simple tasks don't need five stepsLevel-Up Tip
Date and version your workflow docs ("v2 — added a fact-check step, 03 Oct"). The first version is never the final one, and a change log tells you what you already tried.
Action Checklist
○Picked one real recurring task to formalise
○Named its trigger and listed its exact inputs
○Written numbered steps, each with an input and output
○Defined the exact output format and destination
○Marked at least one human checkpoint
Next Guide
22 — Chaining Claude: From Research to Finished Output