• RUBRIC_PATH = rubric_notes.md — Current rubric source
• OUTPUT_DETAIL = full — full: includes per-dimension reasons; compact: score table only

💡 Override when calling:

/cheat-score draft.md — OUTPUT_DETAIL: compact

1. Read

   .cheat-state.json

   → if not exists, prompt user to run

   /cheat-init

   first, then stop
2. Read

   <draft-path>

   → if not exists or empty → throw error and stop
3. Read

   rubric_notes.md

   to find the currently effective formula section (usually in "Current Scoring Dimensions" or "Composite Score Formula" section)

rubric_notes.md

• Current rubric_version
• Dimension list and weights (e.g.,

  ER×1.5 + SR×1.5 + HP×1.5 + QL + NA + AB + SAT

  )
• Normalization constant (e.g.,

  / 8.5 × 2.0

  )
• Meaning of 0-5 scores for each dimension (read from table in "Current Scoring Dimensions" section)

rubric_notes.md

/cheat-score-blind

Spawn cheat-score-blind sub-agent.

Input:
  script_path: <user-provided draft path>
  rubric_notes_path: rubric_notes.md

Task: Score the above script according to the current formula in rubric_notes. Return strict JSON (see cheat-score-blind SKILL.md Phase 2 schema).
Do not read state file / predictions/ / videos/ any other files.
Do not ask the user — you have no user.

• User conversation references / excerpts
• Words like views count / 万 / w / k
• Hints like "Previous prediction was X" / "Actual views were Y"
• Any

  predictions/*.md

  paths

echo "<prompt>" | grep -Ei '播放|阅读|点赞|评论数|实际|retro|复盘|实绩|w$|万$'

1. Parse the dimensions section (includes score + per-dim confidence + reason)
2. Verify

   self_check.any_contamination_signal == false

   , otherwise issue a warning
3. Calculate composite score according to rubric_notes formula (formula logic in main agent, scores from sub-agent)
4. Do not modify the dimension scores given by the sub-agent — scores are for display only. If the user raises an objection (e.g., "AB should be 4 instead of 3"), main Claude records it to

   User Override

   but keeps the sub-agent's original score on file.

refusal != null

• blocked_contaminated_input

  → report that Task prompt contains prohibited fields, ask main Claude to resend

• script_path_invalid

  → check the path

• rubric_unparseable

  → prompt user that rubric_notes.md is damaged

• non_blind_warning

  → still accept dimensions (but all confidence levels are medium), issue a warning

• ❌ Write any files (including predictions/, rubric_notes.md, candidates.md)
• ❌ Provide bucket probability distribution (that's cheat-predict's job)
• ❌ Trigger "published" or "retro" logic
• ❌ Propose rubric upgrades (even if obvious anomalies are found during scoring, only prompt in console, do not modify rubric)

1. Scoring goes through sub-agent. Main Claude no longer does inline scoring. See the isolation protocol in cheat-score-blind/SKILL.md
2. Integer scores only. 4.5, 3.7 are not allowed
3. Blind scoring first. Sub-agent only looks at script + rubric, naturally blind — that's the entire reason for its existence
4. Reasons are diagnostic tools. The 1-30 word reason for each dimension is not decorative — used during retro to identify which dimension's judgment was wrong
5. No file writing. This is the core difference between score and predict. Score is exploration, predict is commitment
6. Do not calculate candidate composite. The composite field in candidates.md is written in cheat-trends/cheat-recommend — score only serves "specific drafted manuscripts"

• "Score and predict at the same time" → Refuse. Please use

  /cheat-predict

  instead. Reason: predict must go through blind check + write immutable logs, score skips these steps
• "Write the score into the observation section of rubric_notes.md after scoring" → Refuse. Observation lifecycle requires that observations must have a "actual performance vs prediction" comparison; scoring alone does not constitute an observation
• "Can you just tell me if it will go viral?" → Refuse. Determining specific composite + bucket requires going through the predict process; score only outputs mechanical calculation under the current rubric
• "Skip the blind sub-agent and let main Claude score directly" → cheat-score does not accept this escape hatch (unlike cheat-predict; cheat-predict has

  --skip-blind

  ). Score is lightweight exploration, no reason to abandon isolation. If Task tool is truly unavailable → prompt user to configure it and try again

• It's a pre-exploration step for

  cheat-predict

  : users can repeatedly score different draft versions, then predict once they've decided on a version
• Score does not update

  .cheat-state.json

  — this is a side-effect-free operation
• If user scores the same draft ≥3 times consecutively → gently prompt in console: "Repeated scoring may introduce decision fatigue, it's time to make a decision"