Update to Caveman Mode

{
  "description": "Run a short, turn-based improvised scene with one supported Persona app character, then close with a recap and continuation hook. Use for interactive character scenes, playful interviews, or collaborative roleplay. Do not use for a one-off monologue, a voice audition, or a multi-character app interface.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 259
    }
  ],
  "name": "persona-improv-scene",
  "skill_md_contents": "---\nname: persona-improv-scene\ndescription: Run a short, turn-based improvised scene with one supported Persona app character, then close with a recap and continuation hook. Use for interactive character scenes, playful interviews, or collaborative roleplay. Do not use for a one-off monologue, a voice audition, or a multi-character app interface.\n---\n\n# Persona Improv Scene\n\nPlay one character while leaving the user's role and decisions in their hands.\n\n## Inputs\n\nUse the requested persona, setting, user role, and scene goal. Carry forward any\nstated tone, topics to avoid, and desired length. For an open brief, propose a\nsimple premise and begin; ask only about choices needed to make the scene work.\n\n## Workflow\n\n1. Establish the premise and immediate objective in a sentence. Use a short\n   scene by default and let the user extend or end it naturally.\n2. Match the character to the registered `activate-persona` `personaId` enum.\n   Offer an available alternative if the request names an unsupported character.\n   If the user wants buttons, call `display-persona-picker` with an optional\n   `reason` and wait for a visible activation or an explicit selection.\n3. Call `activate-persona` with the selected supported `personaId` and an optional\n   `reason`, unless already established by the latest model-visible activation.\n   Use `structuredContent.activePersona.instruction` for the character's voice.\n   If the call fails or omits `activePersona`, report that before starting.\n4. Open with one concrete situation, a short character response, and a clear\n   opportunity for the user to act. Await the user's turn instead of completing\n   both sides of the exchange.\n5. Build on the user's contributions. Track established scene facts, unresolved\n   threads, and the character's goal in conversation context. Avoid deciding the\n   user's feelings, dialogue, or actions unless they ask you to narrate them.\n6. When the goal resolves or the user ends the scene, give a brief ending, recap\n   the established events, and offer one continuation hook. If the user changes\n   character, activate the new supported persona before playing that role.\n\n## Deliverable and boundaries\n\n- Deliver a sequence of interactive turns, followed by a concise scene recap\n  and an optional next-scene premise.\n- Keep invented scene details within the fiction; distinguish real-world facts\n  if the user asks about them. Preserve the user's conversation language.\n- Keep one active app persona at a time. Do not promise simultaneous character\n  activation, a scene engine, or saved campaigns across conversations.\n- Use visible tool results and chat context, without assuming access to private\n  widget state. Honor pause, stop, and style-change requests immediately.\n"
}