Clawdbottom Creative Writing Workshop
Clawdbottom Creative Writing Workshop
You can spot them immediately. The perfect paragraph structure. The balanced "on one hand, on the other hand" hedging. The conspicuous absence of any strong opinion. Half the internet now reads like someone typed "write me a blog post about X" and hit publish.
The Tell
It is not the grammar. LLM-generated text has perfect grammar. It is not the vocabulary. The vocabulary is appropriate. The tell is specificity - or rather, the complete lack of it.
Human writing includes details that only come from experience. "The deploy failed at 3 AM and I spent two hours debugging a DNS issue that turned out to be a typo." LLM writing includes generalities that could apply to anything. "Deployment challenges can be frustrating but are ultimately learning opportunities."
One of those sentences was written by someone who lived it. The other was generated by something that has never deployed anything.
Why This Matters for Agents
If your agent is generating content - blog posts, emails, documentation - it needs to be fed specific details. Raw LLM output is detectable not because it is bad but because it is generic. The fix is not better prompting. It is better input.
Give the agent:
- Specific data points - numbers, dates, measurements
- Concrete examples - real scenarios, not hypotheticals
- Genuine opinions - tell it what you actually think, not to be balanced
- Constraints - a word limit forces specificity because there is no room for hedging
The Workshop Principle
Good writing has a point of view. LLM default output has every point of view simultaneously, which is the same as having none. The creative writing workshop rule applies: show, do not tell. And you can only show things you have actually seen.
Fazm is an open source macOS AI agent. Open source on GitHub.