A couple of weeks ago, I posted about Georgia Appleseed’s pilot study testing ChatGPT, Claude, and Gemini answers to legal questions Georgia parents actually ask. Afterward, a few people asked how I managed the project and if AI helped.

AI did help, but not in the way you might expect. I didn’t use it to just write. I used it more like a Trapper Keeper.

The pilot study had too many moving pieces to manage in a normal chat on ChatGPT or Claude: designing realistic parent questions, legal research, scoring rubrics, blinded review packets, model outputs, spreadsheets, memos, and decisions that needed human judgment. There was so much to pull together and think about.

AI could help with that work. But, I wasn’t sure how to keep it all together and then deliver it back to Claude or ChatGPT.

One evening after dinner, I saw my middle-schooler son pull his papers out of his Trapper Keeper and get started on homework. That’s what I needed: An AI Trapper Keeper.

In the paragraphs below, I describe what an AI Trapper Keeper looks like for me and why it’s a technique that you might try. To make it easy, I share some instructions you can give Claude, ChatGPT, Gemini, Grok, etc., if you’d like to use the technique.

The Basic Idea

To build the Trapper Keeper, I asked Claude to create a “workbook” for the pilot study. (I figured that Claude might take Trapper Keeper a little too literally and get confused.) I used Claude Code, but could have easily used Claude CoWork, or set up a project in Claude Chat or ChatGPT or Gemini. I gave it all the documents and information that I thought would be helpful. I told it to interview me about what I wanted to accomplish and then create a workbook. I told it I wanted a workbook that inventoried what I had and contained a draft study design and plan, a list of additional information I needed, and even a section with questions that I should take to other staff.

I printed it out. And for the next few days, I interviewed staff, thought about the questions, wrote their answers in the workbook, made my edits to the study design in the workbook, and used the lists in the workbook to gather other needed information. Then, I brought it all back to Claude, shared the answers and new info, and told Claude to update the workbook. It incorporated the new info and asked me a few clarifying questions. After a few rounds of this, I had all the information I needed to create a rigorous, doable plan for the pilot study, and all the supporting docs. All that information and the plan itself were part of one, organized document.

Next? I just followed the plan in the final version of the workbook. The last task? Drafting a blog post and getting it up on LinkedIn.

Most people use ChatGPT or Claude one conversation at a time: write a draft, answer a question, summarize a document. A workbook turns AI into a project partner. Instead of solving one task, it helps organize the entire project (the goals, decisions, unknowns, source materials, and next steps) in a single working document.

Each time I worked with Claude, I asked it to update the workbook and create a log. The workbook and log gave me a record of what I learned and what decisions I was making along the way.

Over the last few months, the workbook approach became pretty standard for me. Sometimes a project’s workbook is just a simple document to help me think through writing a speech or creating a presentation or drafting a white paper on AI and child welfare. Other times, it’s much more complex. Using the workbook approach, I did a project that created a simulated focus group and a simulated board of trustees of a foundation, including detailed biographies of each member and rules on how they should interact and make decisions. On a grander scale, I used the method to build a simulation of 1,000 families to see how they might benefit from a particular proposed legislative change.

I even used this method to create a personal curriculum for my own AI Engineering Bootcamp. That bootcamp now automatically produces lessons for me when I start a session, gives me hands-on coding labs, sends me short learning notes each morning, tracks my progress, and picks up where we left off.

Try It Yourself: Sample Instructions

If you have mostly used ChatGPT or Claude by typing a question into a chat box, this is a different way to work. Instead of asking the AI to answer one question or write one thing, you ask it to help you organize the whole project first. Your project could be anything from creating a draft of a document (like a resource list or guide for your clients or a white paper), starting a strategic plan, or even planning your wedding or a home renovation.

Below, I’m posting sample instructions (or a “prompt”) for the AI that you can use to start a workbook session. In its simplest form, the workbook is just a single Word document (or a markdown or html file) with the information I need: what I am trying to make, who it is for, what information I already have, what I still need to find out, who I need to talk to, what decisions to make, and what the final product might look like.

Here is a simple version of the prompt:

I am starting a project about [briefly describe project]. I want you to create a workbook for me that I can print out, complete on my own time, answer any outstanding questions that you have, use to gather information, etc. I'll bring it back to you to discuss further.

Before you draft anything, interview me. Ask me the questions you need answered to understand the goal, audience, source material, constraints, risks, and final product.

Ask one group of questions at a time.

Your AI should ask a lot of questions like: Who is the audience? What is the final product? What tone should it have? What sources are reliable? What facts are uncertain? Who needs to be consulted? What should not be included? What would make the project successful?

After the AI interviews me, I ask it to turn the answers into a workbook. I tell it something like:

Now create a workbook for this project as a single working document.
Include:
- project goal
- audience
- source materials already provided
- information still needed
- people I should interview or consult
- decisions I need to make
- risks or cautions
- proposed outline or structure
- placeholders for anything not yet known

The first version likely won’t be perfect. Sometimes Claude or ChatGPT will misunderstand my audience. Or I will realize that I didn’t give it an important source. It might elevate a small issue that isn’t important. But it might also help me realize that I was ignoring something really important. For example, in my pilot study project, Claude pointed out that my human reviewers, Jack Grote and Tanesse Brown (on our legal services team), needed to have blind scoring packets to avoid bias.

Make edits or tell it what you would like to be different. Work with it until you get the document you need. Then, take your workbook and begin filling it in and collecting any additional info you need.

Use the Completed Workbook to Draft a Document

When the workbook is full enough, I ask the AI to build a plan or draft a document from the workbook.

[Put what you need next in here, e.g., draft the new parent advocacy handbook from this workbook.] Use only the information in the workbook or in the source material it names. If a fact is missing, write [VERIFY]. If a decision is mine to make, write [MICHAEL DECIDE]. Do not guess. An incomplete result is better than an inaccurate result.

Again, the first result will need additional work. Depending on what I’m after, I may have a lot of editing to do. But it’s almost always a great start, because you took the time to gather all the information you needed and organize it. And you’ve saved a ton of time. Check out this flowchart to see how it all works. https://gaappleseed.org/wp-content/uploads/llm-workflow-guide.pdf

That’s it. Once you’ve gotten used to AI asking you questions (instead of the other way around), you can take the next step by creating a permanent memory for your AI (a knowledge workspace, wiki, open brain, etc.). Then things get really interesting. Maybe I’ll share more about some of that in a later post.

If you try the workbook method, let me know how it goes