Make AI use your writing style

How do you get ChatGPT to write in your style? Treat style as a small, reusable spec you turn on when you ask for it. Below is the simplest path I know of to use this technique: define a brief "style card," store a conditional rule, and, if you want, add a one-line self-check. The technique can be used most easily on ChatGPT but it is readily adapted to Gemini and Claude. With a little more effort it can be used on any modern large language model.

The trick is to develop a "style card" that describes the way you like to write. It can be as simple or elaborate as you like. You can write it yourself from scratch or give an LLM samples of your writing style and ask it to abstract stylistic preferences from them. In the latter case, you can progress iteratively, adding elements the AI did not infer or, more importantly, providing negative admonitions of which the AI was unaware because – of course – your writing sample did not contain them. These might include statements like "no boldface except in headings" or "avoid cliches." Here's an example based on the writing style I prefer for this blog. You can have multiple style cards for different forms of writing that you do.

Blog Style:
  Opening paragraph:
    - First sentence states the claim or a crisp definition; total ≤3 sentences.
    - One rhetorical question allowed only if the next sentence immediately answers it and the question is ≤25 words.
    - No meta-promises. A brief italicized abstract is allowed if it contains the claim/definition and nothing else.
  Tone and stance:
    - Formal, direct, evidence-led. Occasional forceful evaluative language is fine when supported by reasons or citations.
  Structure:
    - Headings mark argument turns; paragraphs carry doctrine/analysis.
    - Lists only for steps, checklists, or enumerated factors.
    - Optional scope map after the opening: “We define X, distinguish from Y by mechanism M, evaluate on Z, and state limits L.”
  Mechanisms and examples:
    - After any capability or doctrinal claim, attach at least one of: mechanism, concrete example, metric, or limit.
  Analogies and legal pageantry:
    - Analogies follow mechanism and must lower cognitive load.
    - Avoid legal flattery; prefer low-temperature metaphors.
  Formatting:
    - No inline bold (bold only in headings).
    - Em dashes and parenthetical asides permitted; italics sparingly (including a brief abstract as above).
  Links and citations:
    - Use plain inline links; cite where a reader can verify the decisive step.
  Humor:
    - Dry, restrained, in service of clarity.
  Language constraints:
    - Avoid: paradigm; revolutionize; game-changer; landscape; ecosystem; cornerstone; bedrock; at the heart of; usher in; sea change; genuine reasoning partner; not merely … but …; “This essay will …”.

Store that text somewhere on your file system in case you want to recycle it later.

Now for the key step when using ChatGPT. First, turn on "memory" in your ChatGPT settings. Here's what the settings look like.

And now tell ChatGPT something like the following to ensure that ChatGPT doesn't start writing blog style responses when all you want it to do is describe a good recipe for sous vide peri peri chicken.

What I just gave you is the style card I want you to use when I explicitly ask for writing “in my voice” for essays or blog posts (triggers: “in my voice”, “use my style”, “write as me”, “blog post in my voice”, “essay in my voice”). Add it to "memories." Otherwise use a neutral house style.

And that's really all there is to it. There's one more step you can take to enhance the likelihood that the AI will align its output with your preferences. Add the following admonition.

I want you also to add the following to "memory." After you initially draft a response using blog style, you should check your proposed output against the blog style card. Be particularly attentive to my distaste for inline bold, cliches and contrived legal metaphors. If you find a lack of alignment, revise the response and only then present it to me.

To check whether your request has registered, go directly to the memories ChatGPT stores for you. (Think of it as a Pensieve) (Settings -> Personalization -> Manage Memories). See if it's there. And, if at some point you want to delete the style, just click the remove button. Poof. The memory is gone.

Imposing your voice on other Large Language Models

Gemini

There is very little conceptually different about applying your voice in Gemini. The only difference is that instead of storing the preference in memory, you store it as a "Gem." To create a Gem, open up the "side panel" on the left side of the Gemini web interface. If the menu is collapsed or not visible, you can click on the "Menu" icon, which is typically represented by three horizontal lines (often called a "hamburger menu"), located in the top-left corner of the interface. Click on "Explore Gems." That will give you a list of public Gems (including the fantastic Learning Coach and Storybook) and a "New Gem" button that you should click. You will then see something like this:

Give your Gem a name like "Blog Style." In the instructions field, put in a composite of the three instructions I referenced in the section of this blog entry on ChatGPT: (1) your "Style Card"; a "When to use that Style" command; and, optionally, a "Double Check" admonition. For safety, you can use XML-style tags to delimit these three components. <Style Card>Write like Raymond Chandler.</Style Card><Usage>Respond like Raymond Chandler unless I tell you not to do so for a particular query or some set of queries</Usage><Post Processing>Before you respond to me, check your internal draft and make sure it conforms to the style card. If not, revise it so that it does. Only then give me your response.</Post Processing>. Click the Save button for your Gem. From then on, when you want to have Gemini write in a particular style, go click Explore Gems from the side menu and your own Gem will be waiting.

Claude

With Claude there are two ways to go about imposing your style. Perhaps the simplest is to go to Settings->Profile->Personalization and just plop in the three components, exactly the same way you did as for Gemini. This method will mean that blog style is always available if you want it. You don't have to go into a special mode. The second method is to create a "Project." Although you can do more with Projects than I am discussing here (just as you can do more with Gemini Gems or ChatGPT CustomGPTs), using "Projects" at less than its full capacity appears perfectly fine. Find Projects in Claude's side panel, click on it and then click on New Project. In the resulting web page, give your project a name and – once again – paste in the three components just as you did for Gemini or as I suggested you do it via Personalization. Save. Then, when you want Blog Style (or whatever style you have concocted), just go back into the Project and chat away.

Other Large Language Models

Here's a simple way of using this technique with any large language model. Some of them may have features similar to those I have described for ChatGPT, Gemini, and Claude. But when they don't here is the super-basic backup plan. Remember that file I told you to save containing your instructions? Just call it back up, copy it, and paste it into the start of your chat session. It takes a few more steps than the other methods I have described, but it should be foolproof. To be sure, some models may be better at matching your preferred style than others, but they will all make an effort.

One More Thing

For the next year or two at least, do not expect perfect results. For example, I tried to deploy this technique to write this very blog entry, kind of as proof of the pudding. The result was satisfactory, but not something I wanted to publish unedited. It turns out that I am better at imitating myself than is current AI. (For example, AI would not be able yet to write the preceding sentence – or this one!) That said, the result was not bad and was useful as a first draft. As with many aspects of AI, my advice is to play around. See what works. And keep your hands on the steering wheel. The method I outline here should provide a useful start, but for important work it is not yet a turnkey method.