Skip to content

Digita: a prompt that works like a menu interface, not a single instruction

An open-source MIT-licensed prompt called Digita turns any LLM into a menu-driven interface with reserved keys, state persistence, and portable continuation blocks.

Digita is an open-source prompt (MIT) that turns any instruction-following model into a menu-driven interface. Every response ends in numbered choices — the model never acts until you pick a number.

Three reserved keys are constant in every menu:

  • 8 — view context
  • 9 — update context (restarts the steps)
  • 0 — new task

Task-specific options never collide with those three. Off-script free-text questions get treated as new context, answered, then the menu reprints. Program state survives interruptions.

A visible interaction counter warns when the chat gets heavy and generates a portable “continuation block” — a text snapshot you paste into a fresh chat to resume mid-task without losing state.

NOTE

Digita is pure text. It runs in ChatGPT, Claude, Gemini, or any instruction-following model. No API key, no account beyond the chat interface you already use.

The continuation block is the real innovation here — it sidesteps context window limits by making state portable across chats. The full prompt and source are in the original Reddit thread.