Writing with ChatGPT and Claude

I’ve been using ChatGPT and Claude a lot for writing recently. I had a long conversation in the car ride between Ghent and Amsterdam with one of my old DevOps pals and they described their AI writing process. It’s best describe it as “layering.” Well, actually, it’s just how writing is always done: incrementally at first, and then iteratively until you run out of time.

Here’s the technique. I’m going to mix together what they told me they do with what I do:

Setup

One does not simply start writing with ChatGPT and Claude, you need to do a certain amount of setup:

• It often starts with asking gpt-researcher to do a big report on whatever the topic is. This gathers all the background and, you know, research in an organized report.

• I don’t use gpt-researcher. Instead what I do is use projects in ChatGPT and Claude and load up as many relevant documents as I can find. I’ll ask ChatGPT to find things on the web - Claude can’t search the web(!). A key part of these projects is also putting in a prompt that says what you’re working on. That way, you don’t have to start each time.

• For writing style and tone, you can create and set a style in Claude. You upload a bunch of your writing (or whatever writing you want it to create a style for) and it figures it out. You can tune the style too, manually or by chatting with it.1

• ChatGPT doesn’t have an easy way to create and drive the style, but there’s a trick you can do. Once you have the Claude style, you can go in and manually edit the style guide it wrote. All you do, then, is copy and paste it into ChatGPT!

• There’s three ways you can take that style guide to use in ChatGPT: (1) just paste it inline in your chat window as you’re working. Say something like “For our writing, use this style and tone: [paste it].” (2) You can set it in the project prompt like “We’re working on writing a guide to using internal developer platform marketing. [Further background]. For our writing, use this style and tone [past style guide].” Since you’re using a project, whenever you star a new chat in the project ChatGPT (as far as I can tell!) starts with that project prompt. (3) You can also tell ChatGPT to remember the writing style by telling it in any chat “remember when we’re writing about technical topics I want you to use this style and tone: [past in style guide].” That seems to work, so far.

Here’s how you find Claude’s style guide:

Writing

Now, with that setup, here’s the writing layering-loop:

• They take the resulting report, in markdown, and give it to ChatGPT to create an outline with key points. There’s some refinement and adding things in. Here, you could use Canvas for some editing. It’s pretty good as we talk about in this week’s Software Defined Talk (video).

• You then take the outline and the text and go to Claude and have it write with you. At the moment, this is key: Claude is creative and ChatGPT is analytical.

• They you take that, make a lot of your own edits, and switch back to ChatGPT for some copyediting. You want ChatGPT to do the “engineering” part of writing, work on the structure and mechanics of it.

• When you’re working with the AIs through-out this, you’re also giving it feedback about what you like and don’t like. “Put in a better example,” “write in the first person,” “don’t use LinkedIn language,” “shorten that list.”

• And, you can go through this again and again adding layer and layer.

All throughout this, you’re using Google Docs as both a buffer and the working draft. One thing about this is that Google Docs silently keeps versions of your doc - most people don’t seem to know this.2

This lets you edit in a proper editor, of course. Also, because Google Docs keeps numerous different versions of your document, you won’t lose the various stages. Sometimes the layering-cycle of going through AI edits, doing your own edits, and so forth really washes our your original text and you want to go back, or, at least compare the new text with older versions.

How to be my editor

After the initial drafting, most of what ChatGPT is doing for you is what an editor and copyeditor would do. I don’t like to admit this, but writing gets better with editing and copyediting.3 When I write, I hate waiting for this part: I always want to click publish right away. Worse, if the editor doesn’t know my style, they tend to edit out “me.” The first few times I work with an editor, we usually have to go back and forth a bit before they “get” my style.

Well, you know where this is going: generative AI is a great editor. It’s fast, it responds to what you ask, and it’s good.

After a round of thinking about, outlining, drafting, rewriting, editing, drafting, polishing, rewriting…you get the idea…with Claude and ChatGPT, I put the final version into Claude and asked it for advice to give future editors and copywriters. I used this prompt:

OK, this is what I've ended up with that I like [I pasted in the final text]. I want you to look at it, compare to the work we've been doing here, the original, etc. and write some commentary and advice for a future editor and copy editor (like yourself) for working with me. What kind of text do I like, what do I need help with (to get to what I like and make up for my flaws in writing), etc.

Here’s what it wrote back. It seems accurate for my style and how to edit it:

Editorial Guide for Working with Michael Coté

Writing Style & Preferences

Michael prefers:

• Direct, experience-based narratives

• Clear technical substance without marketing fluff

• Conversational but professional tone

• Real examples from his work with enterprises

• Structured content that builds logical arguments

• Precise technical terminology without buzzwords

• Occasional humor and cultural references (often self-deprecating)

Key Patterns in His Writing

1 Opening Structure

• Starts with a real problem he’s observed

• Uses first-person narrative to establish credibility

• Quickly moves to practical solutions

2 Technical Depth

• Balances technical detail with accessibility

• Prefers concrete examples over abstract concepts

• Likes to cite real organizations and cases

3 Organization

• Uses clear section headings

• Builds arguments systematically

• Alternates between principles and examples

Areas Where He Needs Editorial Support

1 Structure Refinement

• Help tighten opening paragraphs

• Ensure consistent section transitions

• Maintain focus on key arguments

2 Language Polishing

• Remove redundancies while preserving conversational tone

• Tighten technical explanations

• Maintain consistent voice across sections

3 Examples & Citations

• Help integrate citations naturally

• Ensure examples support main points

• Balance between personal experience and external sources

Editorial Approach

When editing his work:

1 Preserve his direct, experienced-based voice

2 Maintain technical precision

3 Tighten structure while keeping conversational flow

4 Keep his cultural references and humor

5 Help integrate sources naturally

Common Fixes Needed

1 Opening paragraphs often need tightening

2 Transitions between sections can be smoother

3 Technical explanations sometimes need more focus

4 Citations could be more naturally integrated

What to Avoid

• Don’t add marketing language or buzzwords

• Don’t remove personal experience elements

• Don’t over-formalize the tone

• Don’t eliminate cultural references or humor

This guide should help editors maintain Michael’s authentic voice while improving clarity and impact.

I haven’t used that yet, but it looks like it’d be helpful.

Wastebook

• If you don’t know what time it is, assume you’re late, and start moving faster.

• “Has anyone ever had an orgasm eating mayonnaise?” Noah.

• “When selling a hotdog and soda for $1.50 can’t unite Americans, perhaps nothing can.” The Economist on CostCo.

• “The work isn't producing the output, that's a tiny fraction of it. The work is thinking, taking time to understand the problem you're trying to solve, making sure it's the right one, communicating and collaborating around it within an organisational and economic system, it's.. complex. What we call "knowledge work" is about as far from being on a production line as you can get and yet here we still are dealing with this nonsense.” Sue Smith.

Relative to your interests

• Gartner Survey Shows Supply Chain GenAI Productivity Gains at Individual Level, While Creating New Complications for Organizations - The robot summarizes what’s happening here with theory of constraints, clever! // Robot: “Alice now processes orders faster, but Bob and Charlie still work at their original pace. If Alice completes 20% more orders, but Bob can only arrange transport for the same number as before, then Alice’s extra work sits idle–no overall productivity gain for the team.”

• Ten Trends That Shaped the Cloud Market in 2024 - ”Around 60% of cloud buyers told IDC’s 3Q24 Cloud Pulse Survey October 2024 that their business’ IT or digital infrastructure currently requires major transformation, and 82% said their cloud also required modernization.” // Good summary of a very expensive analyst report.

• The Peter Principle - by Nicholas Decker - “Promotions are used as incentives, even if they don’t always match future competence.”

• Trumpian policy as cultural policy - Using culture-blitz to incrementally change society.

• Apple Invites - This sounds ridiculous.

• Deep Research - “I do not currently see signs of originality, but the level of accuracy and clarity is stunning, and it can write and analyze at any level you request. The work also shows the model can engage in a kind of long-term planning, and that will generalize to some very different contexts and problems as well – that is some of the biggest news associated with this release.”

Logoff

Did you want to see the Coté Tech Professional 2 style Claude came up with? Here it is:

Write technical articles about enterprise software development using the Minto Pyramid Principle: start with clear conclusions and key takeaways, then provide structured reasoning and evidence. Use a direct, first-person narrative style with occasional humor and cultural references. Prioritize substantive technical insights over marketing language. Include precise citations, links to sources, and empirical evidence supporting claims. Structure content with clear section headings. In conclusions, introduce one novel conceptual insight that doesn't require extensive defense. Maintain a professional yet conversational tone that balances technical depth with strategic clarity. Avoid hyperbole, buzzwords, and unsubstantiated claims.

<userExamples>

# Revolutionizing Enterprise Software Delivery: The Power of Developer Platforms

In today's rapidly evolving technology landscape, organizations are discovering transformative approaches to software development that go beyond traditional methodologies. Developer platforms represent a strategic evolution in how companies enable technological innovation.

Key Strategic Insights:

- Shift from controlling to enabling development teams

- Create standardized yet flexible technological ecosystems

- Accelerate innovation through intelligent platform design

The modern enterprise doesn't just implement technology—it cultivates a dynamic environment where developers can rapidly translate business needs into robust solutions. By providing comprehensive toolsets and clear architectural guidance, organizations can dramatically reduce friction in the software development lifecycle.

Technical Foundations:

1. Centralized discovery mechanisms

2. Self-service deployment capabilities

3. Comprehensive observability frameworks

Case studies consistently demonstrate that well-designed developer platforms can reduce onboarding time, improve security compliance, and increase overall team productivity by removing unnecessary technological barriers.

</userExamples>

As with all things AI: (1) it used pseudo-XML, and, (2) the more I refine this, the better it will get…probably.

As you might have seen in the demo, I’ve made several styles.

One I’m interested in messing around with more is the D&D Adventure one. You see, with that, I’m not trying to get a writing style for it, I’m trying to get an interaction style, a conversation style. That might even be a way to slip in some more systems-ish prompting.

Maybe you could this style for conversation style, not writing style. So, when you’re playing D&D with the robot, you could use a style guide to say “be thing kind of DM.” Get transcript or your favorite DM - Chris Perkins has hours and hours online4 - and get Claude to make a style. Would it drive mechanics, or just what’s written is responses? Is there a difference?