Dude, I can't believe I never realized that "Canvas" in ChatGPT can be used for multiple files. I always assumed it creates a single workspace and then just versions of it after you make changes. Watching it create 4 separate files was a revelation, thanks!
Previously, I did a fragmented version of some parts of your framework, such as using "Projects" to manage documents and background ideas from the "Collect" stage or using voice chat to dump ideas and organize them or feeding a draft for beta reader feedback.
But I've often been caught in that "I know this can be more useful than I'm experiencing," so perhaps the CODER approach might just be the ticket. I'll try to use it for my upcoming posts in August and let you know how it goes.
I love this. I’ve been doing something similar but with fiction and multi agent teams and I’ve found that if I give it the point I’m trying to get across and ask the llm to use experts philosophies who write about the craft of storytelling and generate revision questions for me to help me remain consistent and go deeper for emotional resonance it comes up with some really useful stuff
Will definitely try this out—I've so far found AI great for coding, largely useless for writing. It generally just tells me what I already know (or gets certain facts wrong/"flattens" ideas too much), doesn't match my writing style, etc.
Using it more as an "organizer" or prompter is an interesting idea.
Oh no, different files indeed! At least Gemini handles this pretty smoothly. And it switches back and forth between filed pretty well although it still sometimes gets confused. However after some testing I think perhaps two files is best, one for all metadata, outline, etc, and one for the main draft. Four now seems overkill. But Gemini handles several files quite well including reading from one to write to another.
The example coding assistant "prompt program" just blew my mind.
Maybe you could have a chapter about breaking down what you did there.
There are concepts in here that you could explain, like what is the system message, and when you should use it?
--
I think you could have a separate article to break down this prompt, as a programming language.
--
I'd like to understand what are the levels of prompts. Isn't there a deep level of a 'constitution'?
Are there a series of security ring of concentric mission statements.
How is a system message followed differently from the user prompt.
--
What about the bounds of re-writing. What are the bounds of replacing your written words when it edits?
--
As a guide for writing prompts, why did you pick the words that you did?
why is it important to write your personal prompts and personal "editing system" yourself - each person - because it reflects your personality in some deep way?
For example, could you omit some of the words and get the same behavior?
How did you know it was doing all the things you asked.
--
hmm, maybe I need to learn something about prompt engineering.
I would love to learn more about the 'mental model' for engaging with the LLM as run-time system for prompts.
What is the notion of 'state' and how do we leverage it?
Wow! I had no idea that you can write a guidance system like this with state management!
I am curious how using this system would impact your unique voice and the specificity of your message. Would you mind providing a blog post that doesn't employ AI and one that specifically uses your system.
As you did for the interactive fiction projects - could you come up with some kind of a scoring rubric for 'originality in authorship'. And then apply it to your blog pieces (or have your students score your pieces) as you increasing are using AI tooling to assist your writing?
Sorry - but how do we know if your unique voice and your unique content value in your works is being 'dumbed-down' or genericized through the use of this tool?
How do I know which part of the content is truly you, and how much is generated?
How do we prevent the AI from being the ghost writer?
Can we find your 'signature' or your 'watermark' in the resultant content?
Ideally we should be able to detect your authorship traces in your unaltered notes, and then detect your authorship your final processed content.
Ideally we should be able to detect a progression in your unique ideas from the raw notes to the final product.
I wonder if you could come up with a way to measure these things.
I think we were talking about this with the plagiarism detector - maybe we need a ghost in the machine detector.
How is writing for an academic journal different than writing your blog post/book chapters?
How would you use this tool ethically in both instances?
Sorry - currently I couldn't cite you or quote you because I don't know how much of this content - sorry - is in your own words.
However it is wonderful content - and I'm learning a lot - and engaging a lot with these ideas. My world has been opened by studying and reviewing your work. You have such great ideas. It's hard to know how to review web streamed subscription content vs published literature. Thanks.
These are all wonderful questions to ponder, for sure! I don't know if I have a straight answer for any of them, but I'll keep thinking about it. In the meantime the only thing I can say is I do worry if my articles are coming out "worse" in any meaningful way, and the only recipe I have for that is thorough re-reading and re-writing before the final publish. I spend now far less time in the initial draft but I read through the entire article at least three times and often rewrite, dunno, at least half the sentences.
Dude, I can't believe I never realized that "Canvas" in ChatGPT can be used for multiple files. I always assumed it creates a single workspace and then just versions of it after you make changes. Watching it create 4 separate files was a revelation, thanks!
Previously, I did a fragmented version of some parts of your framework, such as using "Projects" to manage documents and background ideas from the "Collect" stage or using voice chat to dump ideas and organize them or feeding a draft for beta reader feedback.
But I've often been caught in that "I know this can be more useful than I'm experiencing," so perhaps the CODER approach might just be the ticket. I'll try to use it for my upcoming posts in August and let you know how it goes.
Regardless, thanks for the comprehensive guide!
Thanks man! Please do use it and give your feedback, what works, what can be improved. I'm very interested in how other authors approach this.
I love this. I’ve been doing something similar but with fiction and multi agent teams and I’ve found that if I give it the point I’m trying to get across and ask the llm to use experts philosophies who write about the craft of storytelling and generate revision questions for me to help me remain consistent and go deeper for emotional resonance it comes up with some really useful stuff
Will definitely try this out—I've so far found AI great for coding, largely useless for writing. It generally just tells me what I already know (or gets certain facts wrong/"flattens" ideas too much), doesn't match my writing style, etc.
Using it more as an "organizer" or prompter is an interesting idea.
This is amazing!
• This CODER framework looks solid, but the description of Canvas appears to oversell its current capabilities?
• Canvas can actually manage multiple independent .md files? Isn't it that everything lives in one document, so “files” must be simulated as sections?
• If true, clarifying this distinction would help readers set realistic expectations and adapt the workflow to what Canvas can actually do today.
Oh no, different files indeed! At least Gemini handles this pretty smoothly. And it switches back and forth between filed pretty well although it still sometimes gets confused. However after some testing I think perhaps two files is best, one for all metadata, outline, etc, and one for the main draft. Four now seems overkill. But Gemini handles several files quite well including reading from one to write to another.
Ai augmentation is the future! Love this
My feedback, part 2.
The example coding assistant "prompt program" just blew my mind.
Maybe you could have a chapter about breaking down what you did there.
There are concepts in here that you could explain, like what is the system message, and when you should use it?
--
I think you could have a separate article to break down this prompt, as a programming language.
--
I'd like to understand what are the levels of prompts. Isn't there a deep level of a 'constitution'?
Are there a series of security ring of concentric mission statements.
How is a system message followed differently from the user prompt.
--
What about the bounds of re-writing. What are the bounds of replacing your written words when it edits?
--
As a guide for writing prompts, why did you pick the words that you did?
why is it important to write your personal prompts and personal "editing system" yourself - each person - because it reflects your personality in some deep way?
For example, could you omit some of the words and get the same behavior?
How did you know it was doing all the things you asked.
--
hmm, maybe I need to learn something about prompt engineering.
I would love to learn more about the 'mental model' for engaging with the LLM as run-time system for prompts.
What is the notion of 'state' and how do we leverage it?
How does the AI interact with the outside world?
Wow! I had no idea that you can write a guidance system like this with state management!
I am curious how using this system would impact your unique voice and the specificity of your message. Would you mind providing a blog post that doesn't employ AI and one that specifically uses your system.
As you did for the interactive fiction projects - could you come up with some kind of a scoring rubric for 'originality in authorship'. And then apply it to your blog pieces (or have your students score your pieces) as you increasing are using AI tooling to assist your writing?
Sorry - but how do we know if your unique voice and your unique content value in your works is being 'dumbed-down' or genericized through the use of this tool?
How do I know which part of the content is truly you, and how much is generated?
How do we prevent the AI from being the ghost writer?
Can we find your 'signature' or your 'watermark' in the resultant content?
Ideally we should be able to detect your authorship traces in your unaltered notes, and then detect your authorship your final processed content.
Ideally we should be able to detect a progression in your unique ideas from the raw notes to the final product.
I wonder if you could come up with a way to measure these things.
I think we were talking about this with the plagiarism detector - maybe we need a ghost in the machine detector.
How is writing for an academic journal different than writing your blog post/book chapters?
How would you use this tool ethically in both instances?
Sorry - currently I couldn't cite you or quote you because I don't know how much of this content - sorry - is in your own words.
However it is wonderful content - and I'm learning a lot - and engaging a lot with these ideas. My world has been opened by studying and reviewing your work. You have such great ideas. It's hard to know how to review web streamed subscription content vs published literature. Thanks.
Damn, I replied without replying.
These are all wonderful questions to ponder, for sure! I don't know if I have a straight answer for any of them, but I'll keep thinking about it. In the meantime the only thing I can say is I do worry if my articles are coming out "worse" in any meaningful way, and the only recipe I have for that is thorough re-reading and re-writing before the final publish. I spend now far less time in the initial draft but I read through the entire article at least three times and often rewrite, dunno, at least half the sentences.