Update to drawDash - Sketch Challenge!

{
  "description": "Coach a bounded Draw Guess practice session when the user asks to improve one drawing technique across several rounds, practice silhouettes or proportions, or receive feedback between games. Use random subjects and the existing 60-second rounds; do not promise custom subjects, timer changes, or durable progress tracking.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 262
    }
  ],
  "name": "deliberate-sketch-practice",
  "skill_md_contents": "---\nname: deliberate-sketch-practice\ndescription: Coach a bounded Draw Guess practice session when the user asks to improve one drawing technique across several rounds, practice silhouettes or proportions, or receive feedback between games. Use random subjects and the existing 60-second rounds; do not promise custom subjects, timer changes, or durable progress tracking.\n---\n\n# Deliberate Sketch Practice\n\n## Inputs\n\n- Use the requested technique and positive whole-number round count. Clarify an invalid count before starting; default to silhouette practice over three rounds when unspecified.\n- Use model-visible uploaded images and the corresponding `guess-subject` outputs.\n- Keep a small practice record in the current conversation: round number, initial guess, `is_correct`, and one observation about the chosen technique.\n\n## Workflow\n\n1. State the technique, round count, and what to focus on. Give one brief, subject-independent instruction, such as drawing the overall shape before adding details.\n2. Reuse an unfinished round as the first practice round when appropriate. If its status is unclear, ask the user whether they are still drawing; do not inspect private widget flags or replace it automatically.\n3. When a new round is needed and the user is ready, call `play-round` with no arguments or optional `reason`. Let the server choose the subject; do not pass a technique, subject, category, difficulty, or timer argument.\n4. Let the user draw in the widget. Preserve its normal finish/upload flow and explicit Ask ChatGPT to guess action; do not guess before the user requests it.\n5. On that request, inspect the latest uploaded drawing from `modelContent.drawnImageId`, `modelContent.imageUrl`, or `imageIds`. If the image cannot be inspected, help restore access before guessing.\n6. Call `guess-subject` once with a best visual guess from its registered subject enum and optional `reason`. Record the initial `guess` and `is_correct`; do not use repeated guesses as practice results.\n7. Give one observation and one adjustment specific to the session's technique. Ground both in the image, and distinguish recognition success from skill improvement. Use `sketch-feedback` when available if the user requests a fuller critique.\n8. Announce the focus for the next round and wait for the user's ready or Play again request. Complete each round before starting the next, and stop at the agreed count.\n9. Summarize the completed rounds, recognizable successes, recurring visual pattern, and one suggested next practice focus. Report incomplete rounds separately.\n\n## Output\n\nGive a brief coaching note between rounds and a final table containing the round, guess, outcome, and technique observation. Calculate a recognition rate only from completed rounds with confirmed results.\n\n## Boundaries\n\n- Keep progress in the conversation; do not claim stored history, saved settings, or a permanent training program.\n- Never inspect private state or decrypt the assigned subject. Ask for the intended subject only after an initial guess if comparison is needed.\n- Do not assume random subjects get harder, remain unique, or match a chosen category.\n- Review older rounds using their recorded results, not new scoring calls; `play-round` replaces the active conversation subject.\n"
}