Graffiticode and vibe programming with LLMs

This tweet prompted me to write down a few thoughts about how Graffiticode can be used to enable (vibe) programming with LLMs.

Context

LLMs are creative but need strong guardrails for accuracy. Vibe programming works when the program is always working. Then, an author can flow with each iteration until something interesting happens that is worth capturing.

Pattern

Create end-user (and LLM) programmable microapps that are the target for both sophisticated end-users and trained LLMs. They are programmed with task specific language that can be automatically checked for correct semantics as well as syntax. Include a compiler and runtime with error checking that help the AI generate a series of working instances of the microapp for the author to explore / interrogate / play with. Provide a suite of tools for programming the microapp, including an AI helper.

Business Model

  • Build end-user (LLM) programmable microapps with the help of domain experts
  • Open source everything needed to build and program the microapp (code, specs, design docs, etc)
  • Distribute through cloud marketplaces cheaply or free
  • Tune and train LLM assistant to write code for the microapp
  • Monetize through tool (help system) and api (runtime) usage / subscriptions