Start With the Work That Actually Annoys You

Brooks didn’t have his AI aha moment when a tool wrote an email. He already found writing easy.

The breakthrough showed up when he needed to turn a pile of unstructured information into an importable CSV. That job normally meant tedious copying and pasting. He gave the information to AI and received a usable, formatted CSV almost immediately. The value became obvious when the tool removed work he actually disliked doing.

My breakthrough came from turning repeated prompts into a reusable workflow. Start with a real source of friction.

Are you mostly using AI to write slightly better emails? You may be applying it to the wrong work.

Find your RATs first

A fellow workshop instructor shared the filter I use when I look for that friction. Find your RATs before you choose an AI tool. Look for work that is Repetitive, Annoying, or Time-consuming. One category is enough.

Start with a brain dump. Write down the tasks that fit, including the ones you have no idea how to automate. Then ask an assistant which task it could help simplify, speed up, automate, or eliminate.

What on your list is repetitive, annoying, or time-consuming?

You don’t need to know the technical solution before you ask. Brooks recommends starting with the outcome: “I want to do this. Is it possible, and how would I get there?” Describe your constraints. Ask the assistant to walk you through the setup. Curiosity becomes practical when you use the tool to explore its capabilities on a real task.

Pick one RAT and ask what the assistant can do directly, what requires additional access or tools, and what remains manual. Test those suggestions. A confident answer still has to be tried.

If you want a small version of what Brooks did, convert a small, non-sensitive sample of messy information into the format your next system requires. Validate the fields before you import. He described a personal result. Don’t treat it as a speed or accuracy benchmark. Keep the test at that scale.

Make the RAT list this week. Bring it to an assistant. Tackle one item.

Ask five times, then choose

My rule of thumb is to ask at least five times before I settle on a final answer. Five is my heuristic for one meaningful planning task. A simple factual question doesn’t need five turns. More conversation doesn’t replace factual verification.

I inspect the first response. I add missing context. I challenge assumptions. I ask, “What are my blind spots?” I ask, “What haven’t we considered?” The useful answer often emerges through that judgment and those follow-up questions.

Try a five-turn conversation on one planning task you care about. Put the first response next to the final result.

I ask for ten different mock-ups rather than stopping at one or two. I know. Ten sounds like a lot. Ten is my example, and it isn’t an optimal number for every task. AI handles the labor of producing the variations. I supply the taste and the judgment. I identify the elements I like and the elements I dislike across the options. Then I ask for a refined version.

For one visual decision, request a bounded set of distinct directions. Select the strongest elements. Ask for a refined version. Then stop, so the exercise doesn’t turn into endless generation.

The hour after the report

Research that might take days manually can come back as an AI-generated report in roughly 20 minutes. I may spend as much as an hour reviewing it. Those timings are illustrative. They aren’t measured benchmark data. Editing can be faster than authoring from scratch.

Compare the entire manual workflow with AI generation plus verification. Identify who has the expertise to review the result.

Brooks adds a limit that stays unresolved. When you didn’t do the underlying work, you may lack the knowledge needed to judge whether the result is right. Extra review time alone isn’t proof of value. Human approval is only meaningful when the reviewer can evaluate the work.

Can the reviewer actually evaluate the work?

He also warns against sending a slop grenade, a generated message or document that leaves someone else responsible for deciphering it. He remembers a publicly shared AI-generated hiking graphic with wrong distances and nonexistent features. Brooks describes errors and potential danger. Nobody was injured in the account he gave. I’m leaving the person who posted it unnamed.

Before you share anything generated, ask two separate questions. Is this useful to the recipient? Is it actually correct?

Brooks has a hypothetical for software training, and he didn’t report it as a team rollout. “Make a presentation listing the steps” may produce a tidy deck that leaves beginners stuck. Add, “Make sure someone has what they need to use this feature on day one.” That shifts the request toward prerequisites, context, and what happens after the steps. Better instructions define what the reader should be able to do.

Rewrite one AI request around the recipient’s successful next action. Then check the output against that goal.

One workflow, three outputs

I used to upload a transcript, request a summary, request a follow-up email, and separately ask for action items. I explained that same process every time. I configured a workflow that produces all three from the transcript: a summary, a follow-up email draft, and an action-item table. I got the time back because I stopped having to explain the process on every meeting.

An ordinary new chat doesn’t already know these preferences. Configure the workflow if you want all three. The email draft gets a human review before anyone sends it.

Pick one recurring task that feels unnecessarily manual. Put it on your RAT list today. Run that experiment this week. Just one.