Creative ways to talk with your LLMs

LLMs are fun to play around with. Because you can basically instruct an LLM to follow a certain linguistic pattern, it’s just ridiculously fun to have interesting, fun ways of interactions with them. Following are some of my favourites from my sessions in Pi.

Stop dumping text onto me!

I’m reiterating this - LLMs are language models. They have a model of the linguistic features and the vocabulary. They can convey a thing, a notion, an idea, a concept in myriad ways. Think thesaurus on steriods, but that thesaurus can reason about the language mathematically in addition to semantically¹Because the text is trained on human usage of the language rather than the dictionary usage of the vocabulary. This is to say that LLMs can adhere to different ways of representing a textual notion. All you gotta do is, ask it to do so.

And fortunately²As of this writing - Both the frontier labs - Anthropic and OpenAI had released OPUS 5.5 and GPT-6 SOL/LUNA respectively. They’re a stellar improvement on the language articulation from the models it only keeps getting better. That does not mean you can’t experiment with your own style of communication.

Typically, LLMs are eager to dump text onto the user. It’s a hassle to read that - Creates real fatigue and takes the joy out of thinking about the problems/challenges. So try the following in your root AGENTS.md file.³For pi, it’s ~/.pi/agent/AGENTS.md

When explaining a workflow or decision flow during discussion, render a concise Mermaid flowchart directly in the response using a top-level mermaid block. Keep labels short and the graph narrow enough for terminal rendering; do not wrap it in a quote, nest it in another code block, or save it as a repository artifact unless requested. Use diagrams to clarify flow, not to replace necessary decisions or requirements.

Wanna see what this does to your interactions with models in Pi?

A picture is worth a thousand words!