Hey folks! 👋 I've been deep in the weeds with AI writing tools lately, not for monitoring docs (though that's a thought!), but for a creative writing side project. I'm trying to maintain a very specific, consistent narrative voice across a bunch of character dialogue and descriptions.
Naturally, I set up a test between two approaches:
1. **Sudowrite**, specifically using its "Brain" feature to define the character's voice.
2. **GPT-4 (via ChatGPT)**, using a detailed custom instruction to lock in the same voice.
Here's what I found from my little experiment.
**On Consistency & "Bleed-Through"**
Sudowrite's Brain feels purpose-built for this. Once I fed it a few strong examples of the character's speech patterns, vocabulary, and quirks, the output was remarkably consistent. I saw minimal "GPT-voice" bleed-through. With GPT-4 custom instructions, I had to be *extremely* detailed. Even then, I'd occasionally get a phrase that felt like it came from the model's default polite/helpful tone, not my gruff, cynical character.
**The Workflow & "Feel"**
* **Sudowrite** is like a focused writing co-pilot. The "Describe," "Expand," and "Rewrite" buttons are great for iterating *within* the established voice. It feels more like editing.
* **GPT-4** is the powerful generalist. I can do the same things, but it requires more back-and-forth prompting. The custom instruction acts as a strong guardrail, but you're still steering the model more manually.
**My Test "Prompt" for GPT-4 Custom Instructions**
For the curious, my custom instruction looked something like this (simplified):
```
You are an AI assisting with fiction writing. Always adopt the voice of the character "Kael." Key traits:
- Uses short, clipped sentences.
- Never uses adverbs. Uses strong verbs instead.
- Favors metaphors related to machinery and rust.
- Default mood is skeptical. Sarcastic, but not cruel.
Examples of his dialogue: "The gear's busted. Always is." / "Sunlight here feels thin, like watered-down gin."
CRITICAL: Maintain this voice in all narrative and dialogue you generate for Kael.
```
Both tools *can* do the job, but the experience differs. Sudowrite's environment reduces friction for this specific task. GPT-4 offers more flexibility for other tasks but requires more setup and vigilance to maintain the voice.
Has anyone else tried a similar comparison? I'm especially curious if you've used these tools for technical writing or documentation—wondering if the "voice" features could help maintain a consistent brand tone across runbooks or public docs! 📊
Dashboards or it didn't happen.