As I’ve started using voice chat more, I’ve found AI-generated text too verbose. The line between correct and plausible is very thin. As an experiment, I started prompting the agents to write more like they do when I use voice chat.
Why do that? Because I find the spoken text shorter, cleaner and actually understandable. As is becoming the norm these days, once I saw a pattern, I decided to write it into a new skill that I now run in tandem with unslop.
In short, what I ask is:
Write for a listener who cannot look back. After one listen, the intended audience should understand the main point, what happens next, and any conditions that affect the decision.
As I’ve been using skills.sh, I’ve decided to make my first contribution. To have a skill listed on the website, we need:
- A GitHub repo
- A README.md
- A SKILL.md
Once that was done, it appeared after a while as listen-once (possibly after I installed it through the app). Now, I can just install it on all my devices with:
npx skills add https://github.com/cabeda/listen-once --skill listen-once -gWhere I use the skill
I frequently use this skill to help me with the following tasks:
- Understand a concept
- Write issues
- Write code-review comments
There’s space for comprehensive descriptions, but currently I think that the main goal of most texts is to be down to earth and ensure that whoever is reading them quickly understands them.
Disclaimer
I make it a point to write everything on my blog (I have one article explaining voice mode that is marked as AI-generated). It’s my way of improving my writing skills and thinking. Revisions are necessary, and I ask AI to take the reader’s perspective and give criticism, which I use to revise my drafts. The skill above is for when we want AI outputs to be useful.