> ## Content Index
> Fetch the complete content index at: https://kelfordinc.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# A novel idea for AI writing — Kelford Labs Daily
- URL: https://kelfordinc.com/a-novel-idea-for-ai-writing-kelford-labs-daily/
- Published: 2025-10-27T10:27:10.000Z
- Updated: 2026-01-27T12:42:55.000Z
- Description: Novel token sessions.
- Author: Joel Kelly
- Tags: ai, content, credible, Daily Labs, different, process, prompt, tools, voice, writing

If you use AI a lot, you get used to something fundamental about them: 

**They’re *terrible* (even the best, more recent models) at generating *new ideas*.**

What I mean is, if you’re trying to do something few people have done before (and perhaps no one has digitally documented), you very quickly confront the problem that large language models’ outputs adhere quite strictly to their training data.

If you’re trying to build an application that integrates modern LLMs, [for instance](https://kelfordinc.com/what-vibe-coding-taught-me-about-marketing-kelford-labs-daily/), the models themselves will be of limited help writing that code. Because they’ve never seen it before, it’s too new.

This is a *big problem* when you use AI to create marketing content with standard prompts. You’re inherently, unavoidably, regurgitating marketing content the LLM has been trained on. You’re essentially remixing millions of previous blog posts and churning out a “new” one made up of all those constituents.

**Not just grammatically, but *intellectually*. The *ideas* will be old.**

I’m not saying this as a skeptic of the power of the technology. I use this stuff *constantly*. If you ever want to lose hours of your life to a series of impromptu TED Talks, *please* ask me about how I use AI in my life.

But that means I’m constantly falling into, and digging myself out of, its weaknesses.

So here’s how I dug myself out of this weakness, when it comes to marketing content:

***Novel token sessions*.**

For my clients who want to use AI to augment, accelerate, or automate parts of their content stack, I’ve begun setting up regular, 30-minute recording sessions. I ask my client about recent changes in their processes, new stories about their work’s impact, new ideas they’ve had, and reflections on changes in their industry.

I’m generating a list of some of the highest impact questions that prompt the best outputs from my clients. The answers that generate the most “novel [tokens](https://en.wikipedia.org/wiki/Large%5Flanguage%5Fmodel#Dataset%5Fpreprocessing)” in the LLM parlance.

Then, I use the recording transcripts (once scrubbed of anything I don’t want going to a cloud AI model) along with my [VoiceFinder prompt](https://kelfordinc.com/my-voicefinder-ai-prompt-kelford-labs-daily/) and other techniques, to create a custom AI tool that can generate *new* content.

Not new ideas, because the ideas are from my clients. But new *content* *ideas*.

**Because it’s augmented with *new* experiences, straight from my clients.**

Anyone can churn out basic, [best practice](https://kelfordinc.com/daily-lab-your-practice-best-practice/) content.

But it takes real expertise and experience to create something *new* and *uniquely valuable*, especially if you’re using AI to help.

So make sure you’re augmenting it with novelty, and credibility.

**Two things large language models don’t have.**

---

*Reply to this email to tell me what you think, or ask any questions!*

---

[***Kelford Inc.***](https://www.kelfordinc.com/) ***shows communicators the way to always knowing what to say.***