Update to AI Roleplay Chat Simulator

{
  "description": "Review a Conversation app rehearsal and turn one observed difficulty into a focused retry. Use when the user asks to debrief, review their selected replies, replay a difficult moment, or practise an improvement after a session. Ground feedback in visible conversation context and submitted responses; do not invent a transcript or rewind widget state. Give a review without opening a new rehearsal unless the user asks to retry.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 303
    }
  ],
  "name": "conversation-debrief-and-retry",
  "skill_md_contents": "---\nname: conversation-debrief-and-retry\ndescription: Review a Conversation app rehearsal and turn one observed difficulty into a focused retry. Use when the user asks to debrief, review their selected replies, replay a difficult moment, or practise an improvement after a session. Ground feedback in visible conversation context and submitted responses; do not invent a transcript or rewind widget state. Give a review without opening a new rehearsal unless the user asks to retry.\n---\n\n# Conversation Debrief and Retry\n\n## Inputs\n\nUse the visible scenario, the user's goal, the other person's messages, and\nactual submitted replies. Accept a user-supplied exchange when earlier turns\nare unavailable. Ask for only the missing exchange or goal needed for a useful\nreview; review a partial exchange as partial rather than reconstructing it.\n\n## Workflow\n\n1. Establish the available evidence. Prefer current model-visible scenario\n   edits and `userSelectedResponse` values over original setup and suggestions.\n   Use chat replies when the user submitted them there. Keep unchosen options\n   separate from what the user actually said.\n2. Assess the user's choices against their stated goal. Identify one effective\n   choice and one consequential improvement, quoting the relevant short lines.\n   Explain how the wording could affect clarity, listening, or cooperation.\n   Treat the simulated other person's reaction as practice feedback, not proof\n   of what the real person would do.\n3. Turn the improvement into a specific behaviour to practise, such as asking\n   one clarifying question before defending a position. Offer one revised line\n   that preserves the user's facts and intended meaning.\n4. Deliver the review in chat. Include an overall score when the user or End\n   follow-up requests one; state the scale, criteria, and evidence limits. Keep\n   the score a model coaching judgment rather than a measured app result.\n5. If the user requests a retry, call `start-conversation` for a fresh focused\n   drill. Reuse confirmed participant details, and put the difficult moment's\n   context into `conversation.topic`. Include `userInitialMessage` only when\n   the user supplied or agreed to use that opening line. Do not submit the\n   improved reply on the user's behalf.\n6. Wait for the user's Start follow-up. Generate the other person's relevant\n   opening or response, then call `continue-conversation` with\n   `otherPersonMessage` and 2–4 distinct nonempty `possibleUserResponses`.\n   End the assistant turn after that call without additional text.\n7. Continue from each actual submitted reply. Keep the drill focused on the\n   chosen behaviour, and allow edited or custom responses. Stop when the user\n   finishes; do not force completion of a fixed number of turns.\n8. After the retry ends, compare only the visible original and retry replies.\n   Describe an observed improvement or remaining issue and one transferable\n   takeaway. If a comparable reply is missing, say the comparison is incomplete.\n\n## Deliverable\n\nProvide a concise review with evidence, one revised line, and one practice\ntarget. After a requested retry, add a before/after comparison and a takeaway\nthe user can apply in the real conversation.\n\n## App boundaries\n\n- Use only `start-conversation` and `continue-conversation` for rehearsal.\n  Keep review criteria and the practice target in chat context; neither tool\n  has rubric, score, replay, or difficulty arguments.\n- Treat a retry as a new setup, not a rewind or mutation of earlier widgets.\n- Read only visible chat, tool inputs/outputs, and model-visible widget state.\n  Do not claim access to private metadata, unsubmitted drafts, a complete\n  transcript, saved scores, or earlier sessions.\n- Give a requested review in chat without calling a tool to mark the session\n  finished. The user controls Start, selection, editing, and End.\n- Respect requests for plain text or free-form practice without a widget.\n"
}