The difference between people who use AI and people who get hours back is not better prompts. It's that the second group turned three recurring tasks into fixed sequences. Here's how.
Up to now you've been asking for things one at a time. That works. But the real return comes when a task you do every week stops being a conversation and becomes a procedure: the same steps, in the same order, with the same checks, every time. That's a workflow. It's less exciting than it sounds and more valuable than it sounds.
Why one big ask fails
The left side is how most people use AI for recurring work. The right side is what the people who get hours back do instead.
A big vague request makes the model guess at everything: what matters, what format, what to include. It guesses plausibly, which by now you know means generically. And because the output is one big block, checking it is slow, so you either reread it all or trust it all.
Small steps fix both problems. Each step has one job, so the request can be precise. Each output is short, so checking takes seconds. And because the steps are the same every week, you write them once.
Find the candidate
Not everything deserves a workflow. The good candidates share three features:
Recurring. Weekly, daily, every time a certain thing happens. A one-off doesn't pay back the setup.
Same shape each time. Different content, same structure: a report, a reply to a type of enquiry, a summary of a type of document, a first pass at a type of review.
Mostly transformation. Taking material you have and turning it into a different form. Those are the tasks the model is reliable at, because the facts come from your material.
Typical ones: the weekly status report, triaging a shared inbox, summarising meeting transcripts into actions, turning a long document into a one-page brief, drafting the standard reply to a standard request, preparing for a recurring meeting from the last one's notes.
Pick the one that annoys you most. Annoyance is a good signal that the task is repetitive.
Break it down
Write the steps as you'd explain them to a new colleague on their first day. Four or five steps is typical. A useful pattern for most transformation tasks:
Extract. Pull the raw facts out of the material: decisions, numbers, names, questions, dates. "List every…" is the verb. Output: a plain list.
Sort or judge. Group, rank, flag. "Group these by…", "Which of these need a decision?", "Rank by urgency and say why."
Draft. Produce the actual output in the fixed format. The format is pasted in, or saved as a skill (next lesson).
Check. Ask the questions from lesson 3. "Which figures came from my notes?" "What's missing that you'd expect in a report like this?"
Each step is one message. You read each output before the next step. If step 1 missed something, you fix it there, where it's cheap, instead of discovering it in the final draft.
Keep it in one place
Most assistants have a feature for this: Projects in ChatGPT and Claude, Gems in Gemini, or simply a saved note with the steps in it. A project lets you keep the instructions, the format examples and the reference material together, so each week you open it, paste this week's material, and run the steps. It's the "one chat per task" habit from lesson 2, made permanent for a task that recurs.
If your assistant has nothing like that, a text file with the four messages in it, copied in order, works. Unglamorous, effective.
Where the human stays
A workflow isn't automation. You're still the one running it, reading each step, deciding. That's deliberate for now. The model's mistakes cluster in predictable places: things it inferred rather than extracted, numbers it rounded, a category it invented. Your check step catches those, and you'll learn where to look. Later, in lesson 9, you'll meet agents that run steps on their own. The workflow you build now is exactly what you'd hand one, and you'll know where it needs supervision because you've done it by hand.
Measure it, once
After the second or third run, time it. Honestly, including your checking. Compare with how long it used to take. Most people find a task of an hour becomes twenty minutes, with the twenty minutes being mostly reading and judgement rather than typing. If it didn't get faster, the steps are wrong, usually too big. Split the slow one.
The habit
One recurring task, broken into steps, written down, run every time the task comes round. After a month, a second one. That's the whole method. People who do this for three tasks don't talk about "using AI"; they just have shorter Fridays.
Try this today
Pick the recurring task that annoys you most. Write its steps in a note, four or five of them, with the exact message you'll send at each step. Run it once on last week's material, as a rehearsal. Fix the steps that produced something you had to rewrite. Next time the task comes round, run it for real.
Check yourself
5 questions. Everything is in the lesson above. Get them all right and the lesson counts as done.
1. Why does "write my weekly report from these notes" usually produce a report you have to rewrite?
One big ask means many guesses. Small steps let each request be precise and each output short enough to check.
2. Which task is the best candidate for a workflow?
Recurring, same shape each time, mostly transformation of material you already have. The other two are one-offs or judgement calls.
3. In the extract-sort-draft-check pattern, what is the point of reading each step's output before the next?
A missed item in the extract step is a one-line fix there. In the finished report it's a rewrite.
4. Where should the steps and format for a recurring task live?
A project keeps instructions and format together without the crowded-window problem. A long chat is the opposite of that.
5. After three runs the workflow is no faster than doing it yourself. Most likely fix?
Steps that are too big reintroduce the one-big-ask problem. Find the one you rewrite most and break it in two.