8 Comments
User's avatar
Annette Vee's avatar

This is a great tool! I especially appreciate the "potential leads" section and think that might be useful for teaching students about research projects. This was my own attempt, testing a claim on the 5 paragraph essay: https://chatgpt.com/share/6852e325-e954-800e-bf41-c8a2d2623f94 Thank you for sharing!

Expand full comment
Arnav B. Dhanuka's avatar

Thank you for bringing the toolbox to ChatGPT. I know cutting your prompt down by two-thirds must have been tough. Losing the part where GPT analyzes photos isn’t ideal, but as you mentioned in the blog post, the toolbox is now available to a much wider audience. That’s a big plus.

Hopefully you’ll make another toolbox that focuses on image analysis. In my experience, giving ChatGPT a specific task or even a smaller subtask often helps it work better.

Expand full comment
Arnav B. Dhanuka's avatar

Hi Mike,

After a few days with this GPT, one thing is obvious: the “another round” prompt is overloaded. It’s used both to double-check answers and to dig deeper. That overlap creates confusion.

We should split the functions. Use “go deeper” for further exploration and introduce a separate prompt like “double-check” for accuracy. Clear prompts lead to better results.

You’ll need to modify the GPT instructions accordingly. Separate these functions so that each prompt does one thing well.

Expand full comment
Rich James's avatar

Bravo! Sharing widely.

Side note, my company encourages using AI but blocks use of external GPTs. Likely because we have an enterprise version to protect client data. Likely others in the same boat. I go to Claude to use SIFT and avoid client data. I develop learning materials and use it understand unfamiliar subjects and check the evidence for the assumptions of my profession.

Expand full comment
Anna Mills's avatar

This is so exciting!

Expand full comment
Alberto Balsam's avatar

Have you experimented with using more terse language for the toolbox prompt, and seeing if performance is maintained? For example if you ask an LLM to rewrite in "telegraphic style" you'll get all the semantic content maintained but at about 50% of the size, albeit harder to parse (for a human) due to the clipped syntax.

Expand full comment
Mike Caulfield's avatar

I might later, but I've just had to throw away almost 40% of the prompt so I am still mourning

Expand full comment
Dmitry Erkin's avatar

You don’t have to throw it away!

Add full instructions as a document to your custom chatgpt

In your prompt just write something like “You are … follow the instructions in doc name”

LMK if you need help with it!

Expand full comment