Update to 1 Billion Brain Cells

{
  "description": "Compare completed attempts at the same Brain Training game using results visible in the conversation or supplied by the user. Use for progress reviews, replay comparisons, or questions about changes in scores and errors; do not use to rank different games, retrieve unavailable history, or interpret one round alone.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 274
    }
  ],
  "name": "compare-brain-training-rounds",
  "skill_md_contents": "---\nname: compare-brain-training-rounds\ndescription: Compare completed attempts at the same Brain Training game using results visible in the conversation or supplied by the user. Use for progress reviews, replay comparisons, or questions about changes in scores and errors; do not use to rank different games, retrieve unavailable history, or interpret one round alone.\n---\n\n# Compare Brain Training Rounds\n\n## Inputs and workflow\n\n1. Identify the game and collect at least two distinct completed attempts from\n   model-visible results, completion follow-ups, or user-supplied records. Ask\n   for an earlier result if only one is available. Do not call a game-launch\n   tool to fetch history; the app has no result-history tool.\n2. Group records by the exact registered game: `color-switch`, `number-sort`,\n   `odd-one-out`, `path-trace`, `pattern-recall`, or `word-run`. Compare within\n   each game. Use visible timestamps or the conversation's stated order; do\n   not invent dates, fill missing fields with zero, or count a remounted widget\n   as a new attempt. Preserve user-reported data as such.\n3. Read the success/failure field pair from the result contract: prompt counts\n   for Color Switch, word submission counts for Word Run, and round counts for\n   the other games. Use the reported score where present. Number Sort results\n   store counts rather than a score, so derive `(successfulRounds - failedRounds)\n   * 100` if needed; Odd One Out uses `successfulRounds * 100`. Do not reconstruct\n   Word Run scores from counts because accepted word lengths affect the score.\n   Check the available tool contract for other missing or conflicting scores.\n4. Calculate changes in score, successes, and failures wherever both attempts\n   supply those values. Calculate success share as `successes / (successes +\n   failures)` only with both counts and a nonzero denominator; show its change\n   in percentage points. Label Color Switch failures as combined wrong or\n   expired prompts, not wrong answers alone. Prefer absolute score changes to\n   percentage changes when the baseline is zero or negative.\n5. Explain whether the visible change reflects more successes, fewer failures,\n   or a trade-off between them. For Word Run, higher score with similar counts\n   can reflect longer accepted words, but do not identify particular words or\n   causes without evidence. For Odd One Out, discuss failures separately from\n   the score because they do not subtract points.\n6. Acknowledge variation from randomized content and, in the memory games,\n   increasing sequence or route lengths within a session. Describe an observed\n   difference rather than proving a lasting change in cognitive ability or a\n   particular strategy's effectiveness. End with one measurable next-attempt\n   target based on the user's goal and the known results.\n\n## Deliverable and boundaries\n\nReturn a small comparison table with attempt label, game, known score, successes,\nfailures, and success share. Mark unavailable values as unknown, add relevant\nabsolute changes, and give a short interpretation plus one next target.\n\nUse a score-only comparison when counts are unavailable, stating what remains\nuncertain. Do not infer reaction times, unfinished attempts, individual mistakes,\npopulation benchmarks, or private widget state. Do not collapse scores from\ndifferent games into a leaderboard. Keep this workflow analytical; launch a\nfresh matching game only if the user also requests another attempt, with at most\nan optional string `reason`.\n"
}