Lessons from building an LLM-first framework

We’ve been developing Tonk for shareable, multiplayer applets enriched with context and intelligent from across your life. What makes developing these applets a little different is that they are “malleable”, that is, they are meant to be updated quickly by the users of the application through vibecoding. At first, we struggled to keep the coding agents from going off the rails, but now our framework, to our surprise, is totally usable by non-coders.

I’ll talk about the two ways we achieve this:
1) eliminating the need for context beyond the frontend (ie. what the agent can see)
2) prompting with lots and lots of hint files

I can also talk about a few failed experiments: using a recursive task definition framework for the agent and injecting context through MCP plugins.