cheat-shoot

/cheat-shoot — register shooting complete + create video folder + (if script changed) trigger v2 prediction

Advance a video from the "prediction written, not shot" state to the "shot, not published" state. This step:

Create videos/<同 id>/ directory (if not already present)
Ask the user: "Is the script used during actual shooting consistent with scripts/<id>.md ?"
Compute diff — exceeds V2_TRIGGER_THRESHOLD (default 30%) → delegate to /cheat-predict — mode: v2 append to the original prediction file ## 预测 v2 section
Add the video folder to the state.shoots queue, buffer +1

cheat-shoot itselfdoes notwrite prediction content — all prediction persistence logic is in cheat-predict. cheat-shoot is only responsible for detecting script changes + dispatching.

Why a separate skill:

The buffer warning system needs to clearly distinguish "shot" vs "published". Videos can be shot in batches (5 in a day), published spread out (1 per day)
"Actual shooting script" ≠ "pre-shoot draft" is the norm. This step is the entry point for making the diff explicit, triggering v2 re-judgment, and collecting "user script-change pattern" signals
The difference between v2 prediction and v1 prediction is itself evidence of rubric upgrading — e.g., v1 gives ER=4, v2 gives ER=5 (the user's script change increased hook strength), which tells the rubric "this user's ER threshold is inconsistent with my current formula"

Overview

[用户：拍了 scripts/2026-05-04_abc123_停止期待.md]
  ↓
[Phase 0: 解析路径 + 验证 prediction 已存在]
  ↓
[Phase 1: 检查是否已登记（避免重复）]
  ↓
[Phase 2: 建 videos/<id>/ + 询问"实际拍摄稿一致吗？"]
  ↓
[Phase 3: 写 videos/<id>/script.md]
  ↓
[Phase 4: append state.shoots]
  ↓
[Phase 5: 输出 buffer 状态]

Constants

REQUIRE_PREDICTION = true — a v1 prediction file must exist before shooting
V2_TRIGGER_THRESHOLD = 0.30 — after normalization, char-level diff exceeds 30% → by default recommend v2 re-judgment; below 30%, ask the user whether v2 is still wanted
DIFF_METRIC = char_levenshtein_normalized（default)—— via tools/diff_pct.py call: first normalize (remove markdown headers / separators / list markers / decorative punctuation / collapse all whitespace), then compute char-level Levenshtein / max(len_a, len_b). preferred backend rapidfuzz，fallback difflib.SequenceMatcher(stdlib, always available).The old line-level version produced severe false positives in conversational transcription scenarios(draft has long markdown sentences vs whisper transcription's short broken sentences, content almost unchanged but line-level computes ~200% diff) — fixed in PR #14
DIFF_METRIC=lines —— legacy fallback: when python3 is completely unavailable or tools/diff_pct.py cannot be found, degrade to diff -u | grep '^[+-]' | wc -l algorithm

Inputs

required	source
`<scripts-path-or-id>`	user parameter; if missing, ask
`.cheat-state.json`	state file
`scripts/*.md`	pre-shoot draft
`predictions/*.md`	verify the corresponding prediction exists

Workflow

Phase 0: Parse + validate

Parse the path given by the user — several forms are supported:
- full path scripts/2026-05-04_abc123_停止期待.md
- abbreviated 2026-05-04_abc123_停止期待
- id abbreviation abc123 → glob scripts/*_abc123_*.md find a match
Verify scripts/<id>.md exists: if not → error "cannot find pre-shoot draft"
Verify a corresponding prediction exists predictions/<同名>.md：
- does not exist → refuse registration, prompt "run /cheat-predict first to write a prediction, otherwise it violates the blind prediction principle — you cannot write the prediction only after shooting, that amounts to writing it after seeing the footage"
- Exists → Pass

Phase 1: Check for duplicates

Read .cheat-state.json, check whether shoots[] already contains this id:

Already exists → Warn "Already registered (X days ago). Do you want to re-register, or publish with /cheat-publish?"
Does not exist → Proceed to Phase 2

Phase 2: Create video folder + ask about script consistency

Create directory videos/<id>_<short>/(same naming as scripts/ + predictions/)
Ask the user：

拍 「<title>」 的时候，你实际用的稿子和 scripts/<id>.md 一致吗？

a) 一致——按草稿拍的
b) 改了一些——你能给我看下实际拍摄稿吗？我重新打分一次（v2 预测）
c) 大改了，基本是另一条 → 走 _redo 流程：
   scripts/<id>_redo.md → 重新 cheat-predict → 再 cheat-shoot（原 prediction 留档脱钩）

Phase 3: Write videos//script.md + (b path) trigger v2 prediction

a path (consistent)：

cp scripts/<id>.md → videos/<id>/script.md
script_consistency = consistent
Do not re-judge, proceed to Phase 4

b path (changed)：

Ask the user for the actual shooting script—pasted text / file path / transcript file
If the user provides it → write to videos/<id>/script.md
If the user did not keep it (improvised) → mark script_lost, write a placeholder file + warn "v2 re-judgment skipped—next time it is recommended to keep a script (even a voice memo transcript)", proceed to Phase 4

If provided: compute diff

# 解析 cheat-on-content 源码根（cheat-shoot 是 symlink 装的）
SKILL_REAL="$(readlink -f ~/.claude/skills/cheat-shoot 2>/dev/null || readlink ~/.claude/skills/cheat-shoot 2>/dev/null)"
if [[ -n "$SKILL_REAL" ]]; then
  REPO_ROOT="$(cd "$SKILL_REAL/../.." && pwd)"
  DIFF_TOOL="$REPO_ROOT/tools/diff_pct.py"
fi

if [[ -n "${DIFF_TOOL:-}" && -f "$DIFF_TOOL" ]] && command -v python3 >/dev/null 2>&1; then
  # 默认 char-level Levenshtein on normalized text（rapidfuzz preferred, difflib fallback）
  diff_pct=$(python3 "$DIFF_TOOL" "scripts/<id>.md" "videos/<id>/script.md")
else
  # legacy line-level fallback——只在 python3 或 diff_pct.py 都不可用时用
  added=$(diff -u scripts/<id>.md videos/<id>/script.md | grep -c '^+')
  removed=$(diff -u scripts/<id>.md videos/<id>/script.md | grep -c '^-')
  total_orig=$(wc -l < scripts/<id>.md)
  diff_pct=$(( (added + removed) * 100 / total_orig ))
  echo "⚠️  fallback 到 line-level diff——口语化转录会 inflate diff_pct，可能误触发 v2"
fi

Why normalize + char-level: line-level diff in creators' real scenarios (the draft is long markdown sentences, the shooting script is short colloquial lines from a whisper transcript) computes ~200% difference but the content is almost unchanged. After normalize, char-level Levenshtein stably reflectscontentdifferences, rather than formatting differences. See tools/diff_pct.py + tools/diff_pct_test.shfor details (3 fixtures all pass on both backends).

Determine v2 trigger：
- diff_pct >= 30 → by default recommend v2 re-judgment,actively call /cheat-predict — mode: v2 — prediction-file: predictions/<id>.md pass videos/<id>/script.md as input. cheat-predict runs in v2 mode and appends ## 预测 v2
- diff_pct < 30 → ask the user: "Only N% of the content was changed, do you want to re-judge? Default is no (the v1 prediction is still valid)". If the user says yes → call as above; if the user says no → skip v2, continue to Phase 4
After cheat-predict completes the v2 write to disk, control returns to cheat-shoot to proceed to Phase 4

c path (major change)：

Do not write videos/<id>/script.md, prompt to follow the _redo flow
Exit cheat-shoot (do not proceed to Phase 4)

Phase 4: state update

{
  "shoots": [
    ...,
    {
      "video_folder": "videos/2026-05-04_abc123_停止期待/",
      "prediction_file": "predictions/2026-05-04_abc123_停止期待.md",
      "scripts_path": "scripts/2026-05-04_abc123_停止期待.md",
      "shot_at": "<ISO timestamp>",
      "script_consistency": "consistent" | "modified" | "lost",
      "script_diff_pct": <0-100 int 或 null>,
      "v2_prediction_written": <true/false>,
      "script_hash_at_shoot": "<sha256:12 of videos/<id>/script.md>"
    }
  ]
}

v2_prediction_written: true indicates that the prediction file now has the ## 预测 v2 section; cheat-retro should read v2 to compute deviation;false indicates to continue using v1.

Phase 5: Output buffer status

Immediately after reading state, compute buffer + color (per cadence-protocol.md 's derivation rules):

✅ 已登记拍摄：videos/2026-05-04_abc123_停止期待/
   预测文件：predictions/2026-05-04_abc123_停止期待.md

📦 当前 buffer：3 篇（🟢 绿色，正常）
   按你的 cadence（隔日更）= 6 天 buffer，节奏稳定。

下一步：拍其他候选 / 等下个发布日 / 不动

If the buffer color changed (e.g., from green to blue) → highlight reminder:

📦 当前 buffer：6 篇（🔵 蓝色，**积压**）
⚠️  建议暂停拍摄，全力发布存货 + 复盘。
   按你的 cadence（日更）= 6 天预备，已超过健康上限。

Key Rules

Do not write prediction—shooting ≠ publishing. The prediction is locked at /cheat-predict, shooting is just an event
Do not touch video folder contents—neither script.md nor draft-v0.md is changed
There must be a prediction first—otherwise it violates blind prediction (looking at the footage after shooting and then writing a prediction = data leakage into judgment)
Compute buffer in real time—recompute immediately after each shoot / publish; state.shoots is the ground truth
Batch supported: the user can say "shot X / shot Y / shot Z" three times in a row in one day and register them consecutively

Refusals

"Shot X, but I never ran cheat-predict" → Reject. The v1 predictionmust be written before shooting——Writing a prediction after shooting will be influenced by the footage and lead to after-the-fact edits. Please run /cheat-predict to write v1 first, then come back to /cheat-shoot. (Re-judging as v2 is a separate matter—only allowed if v1 already exists + the draft is revised after shooting)
"I don't have a video folder, I shoot directly" → ask the user → help them create a video folder + remind them to follow the full process next time; when registering, mark ad_hoc: true
"I revised the draft but just overwrite v1, don't keep a v2 section" → refuse. v1 is the archive, v2 is the current judgment—append, don't overwrite. Keeping both sections together is key evidence for rubric learning

Integration

Upstream:/cheat-predict finish writing prediction → user shoots → /cheat-shoot register
Downstream:/cheat-publish when publishing, remove the corresponding item from state.shoots
/cheat-status the dashboard's buffer number comes directly from state.shoots.length
/cheat-recommend adjust the recommendation strategy based on the buffer color
The SessionStart hook looks at the buffer color to decide the first line of the report

state.shoots data structure

{
  "shoots": [
    {
      "video_folder": "videos/2026-05-04_abc123_停止期待/",
      "prediction_file": "predictions/2026-05-04_abc123_停止期待.md",
      "shot_at": "2026-05-04T18:30:00+08:00",
      "ad_hoc": false  // true if user shot without going through full flow
    }
  ]
}

In ascending order by shot_at —the earliest shot comes first./cheat-status Show a days-since-shoot warning for the earliest item (to avoid having a video that was shot 30 days ago and still not published).