Find (and Fix) Your AI Blind Spots

How long have you been using AI now?

If you’ve been using OpenAI’s ChatGPT, Anthropic’s Claude, or Google’s Gemini for a year or more, you’ve probably settled into a rhythm.

You know the handful of things it’s good at. You have your go-to tasks. And, as the models improve, some of the outputs you’re getting for those go-to-tasks get even better. But, somewhere in the back of your mind, you suspect you’re not quite using AI to its full potential.

It’s hard to avoid feeling a tinge of FOMO (fear of missing out) when your social feeds are constantly surfacing new AI use cases other people are finding success with.

Inspiration from others can play a big part in discovering new ways to use AI. But there’s a more effective way to find high-signal advice on how to improve your AI usage -- and it comes directly from the AI you’re already using.

So today I want to break down:

• What is meta prompting and how has it changed?

• How to ask your AI for a usage audit

• How to continuously learn firsthand from your AI

—@dharmesh

Meta Prompting: What’s Changed?

Last year around this time, I devoted a whole newsletter to a technique called meta prompting: instead of struggling to write the perfect prompt yourself, you ask the AI to write a prompt for you, giving it permission to ask clarifying questions along the way.

I even built a free tool called MetaPrompt.com that does exactly this. You type in the prompt you regularly use, and it gives you a better one.

That was ten months ago. The model I recommended in that issue has been superseded twice in the past three months alone. Since then, meta prompting didn't age out -- it kept growing.

Back then, meta prompting was about careful wording. You brought a draft prompt, the AI tightened it, and its clarifying questions caught the details you forgot to mention. It also doubled as a learning tool: watching the AI restructure your rough asks taught you what a good prompt contains, and over time you wrote better first drafts on your own.

The newer version of metaprompting practiced today has you interviewed by the AI before anything gets written. Here’s an example prompt for you to try it yourself:

You don't know everything the AI needs to know to do the job well -- but it does. Asking it to interview along these lines is what helps cover your blind spots. The added context that comes from the interview can make a big difference in the quality of your output.

Power users (naturally) have pushed further. A recent tip from OpenAI’s Vaibhav Srivastav made the rounds on X with roughly 178,000 views:

If that tweet reads like another language, don't worry. Let’s break it down:

• Codex is a harness made by OpenAI.

• Goals are the native way to create loops inside of Codex.

• Vaibhav is outsourcing the work behind loop creation back to the model itself to great success. Another meta prompting win!

This shows how far frontier models have come since we first discussed meta prompting. We’ve reached a point where our long, multi-paragraph prompts have become “Go research this thing, determine an appropriate goal based on your research, then go work on it until the goal is complete”.

Here are a handful of simpler prompts you can use to apply the same concept:

• "Before you start, tell me what context you need from me to do this well."

• “Restate what you think I'm asking for before you act."

• “Ask me questions one at a time until you’re confident you understand the job, then start.”

Try one yourself (or better yet -- try them all!) to set your next task up for success. The concept to take away here is that the AI is remarkably good at identifying what it doesn’t know. You just need to give it permission to surface the question marks.

Ask For a Usage Audit

Another strong way to get better at using AI is to conduct a usage audit.

This involves exporting all your chat history, handing it back to AI, and asking it to identify trends, optimization opportunities, and all kinds of other fun tidbits.

There are several built-in ways to call for a usage audit. Because these differ depending on which model you use most and how your memory settings are configured, we’ll take the manual approach. This takes a bit more elbow grease but will work for everybody.

Step 1: Export Your Data

First off, we need to save a copy of all our data. These exports will include all your conversations and projects, which is exactly what we need to perform a usage audit.

Claude users:

• Click your name in the bottom-left corner

• Settings → Privacy

• Under the “Your data” heading, click “Export Data”

ChatGPT users:

• Click the icon in the bottom-left corner

• Settings → Data Controls

• Find “Export data” and click the “Export” button

Oh, and in the case the details of how to do this have changed since the time I write this (AI moves fast!), you can do what I do: Ask Perplexity.ai how to do it. It works really well at that particular use case (telling you how to use a piece of software).

Both Claude and ChatGPT warn that it could take a day or so to receive the download link for your data. It gets easy from this point forward, so don’t worry if you need to come back and revisit things tomorrow when the download is ready.

Step 2: Analyze Your Data

Your data will download as a ZIP file. Open that up and you’ll find conversations.json. To analyze every single conversation you’ve ever had with AI, all you need to do is drag-and-drop this file into a new conversation and pair it with a prompt.

You can get creative here -- don’t be shy! Here are two example prompts you may find useful in analyzing your usage:

• Paste this prompt: “What tasks do I often ask you for repeated help with, and how could we get a better result from those tasks?”

• Paste this prompt: “Find the conversations where I got frustrated, gave up, or ended the conversation before the desired result was achieved. What went wrong, and how should I have prompted differently?”

The idea is to have a casual conversation that surfaces where you can work together more efficiently. The AI has all the data it needs thanks to the file you supplied. If you run out of questions to ask, pass the baton back to AI by asking: “What are the 10 most important insights you can glean from my AI usage to date?” 

Learn AI, From AI

Your social feeds can show you what tactics are working for everyone else. The AI you’ve worked with is the only one that has built up the firsthand context to see what’s worked (and what hasn’t) in your actual workflows.

Leaning on AI to use AI better is a powerful meta skill that can increase your rate of learning well beyond what external learning materials alone can do.

And none of this is a one-time exercise. Meta prompting to address the unknowables is a skill you can practice with every project. Analyzing your AI usage (with AI) feels like a worthy use of your time every month or two.

Now, I want to ask you something.

In the spirit of getting better at leveraging AI, I have a question for you: what’s the most useful AI tip you’ve implemented in the past 3-6 months?

Hit reply and let me know your answer. The responses I get to this newsletter are a gift -- I read every reply, and respond to a few where I can.

—Dharmesh (@dharmesh)