• DEFAULT_RETRO_WINDOW_DAYS = 3
• INSTALL_HOOKS = ask — Ask by default; users select

  auto

  to install directly;

  skip

  to skip installation
• TREND_DEFAULT_SOURCES = ["manual-paste"]

1. Read the user's current working directory (the user's content project, not cheat-on-content itself)
2. Check if

   .cheat-state.json

   already exists:
   • Exists → Prompt: "The project seems to have been initialized (state file exists). Reinitializing will overwrite existing configurations — confirm?" Only proceed after the user explicitly confirms
   • Does not exist → Proceed to Phase 1
3. Check if core files such as

   rubric_notes.md

   /

   predictions/

   exist — if they exist but the state file does not → it's a "semi-initialized" state, prompt the user and ask: "Do you want to infer the state from existing files or reset?"

"Which of the following does your content resemble more? a) Opinion video (commentary / current affairs commentary / argumentation / topic discussion / personal opinion) — directly matches the built-in rubric b) Long essay (WeChat Official Account / Substack / Medium) — can start with the opinion video rubric, adjust weights during bump c) Short text / thread (X / Weibo / Jike) — same as above d) Podcast / long-form video (YouTube long video / podcast) — same as above e) Tutorial / tool teaching / Builder (teaching others how to use X tool / how to do Y project) — same as above f) Other (gaming / food / makeup tutorial / news / drama) — workflow is universal, but rubric dimensions need adjustment (The ER/SR/HP set may not be predictive for your form, you need to break down suitable dimensions yourself) g) Mixed"

content_form

rubric_form_mismatch

content_form

content_form

"opinion-video"

"long-essay"

"short-text"

"podcast"

"tutorial-builder"

"other"

"mixed"

rubric_form_mismatch

• Select a →

  false

• Select b/c/d/e/f/g →

  true

  , cheat-status will continuously prompt "Your content form may require bump weight adjustment"
• No more "severe mismatch" category — all forms can run the workflow, only some rubrics require more aggressive bump

"What's the typical duration of your videos? a) 30 seconds - 1 minute b) 1-3 minutes c) 3-5 minutes (recommended for getting started) d) 5-10 minutes e) Over 10 minutes"

typical_duration_seconds

"How often do you plan to publish? a) Daily b) Every other day c) Weekly d) Flexible / irregular (disable buffer monitoring)"

target_publish_cadence_days

"a) No — I will help you brainstorm 5 candidates based on interests + hot topics + write 5 drafts b) Yes — whether it's 1 or 100 videos, I will help you fetch historical data to make subsequent brainstorming more aligned with what you've done"

calibration_samples: 0

"Which platform is your content mainly on? a) Douyin — install douyin-session adapter (Playwright + scan QR code to log in to Douyin Creator Center) b) YouTube — install youtube-data-api adapter (requires API key) c) Bilibili — bilibili-stat adapter d) Other / multiple platforms — use manual paste mode"

"Install the adapter now for automatic fetching, or tell me manually first?

• Install now — guide you to install Playwright + scan QR code → fetch the latest N pieces of data
• Install later — use manual mode first, mark state as 'pending_adapter_setup', cheat-status will continuously prompt to install"

"How many of your latest videos can I fetch as the basis? (Recommended 10-25 videos; the more samples, the more accurate the baseline. Up to the actual number of videos on your account)" → User provides number N, fetch N videos in Phase 3.5

"Approximately how many videos have you posted? Just give a range (e.g., '5-10 videos' / '20+ videos'), This is only used to mark the estimated value of calibration_samples, no need to be accurate." → User provides an estimated value, skip fetching in Phase 3.5, write the estimated value to calibration_samples

"How to retrieve data for T+3 day retrospective?

a) Manual paste — backup option. You must paste the top 20+ comments (with like counts), not just view counts. Comments are the real signal — meme explosions like 'she's different' can only be seen from comments, View counts will never tell you what content truly resonates with the audience. b) [Recommended default] Adapter automatic fetch — get all comments + data. If you haven't installed the adapter yet, it's okay, mark state as 'pending_adapter_setup', Install it before the first publish (cheat-status will continuously remind you to install). Installation guide is in adapters/perf-data/<platform>/README.md."

data_collection

data_collection

"manual"

"adapter"

"Do you have a list of candidate topics now? (e.g., maintained in external markdown / Notion) a) No (default) — I will help you brainstorm later, or use /cheat-trends to fetch topics daily b) Yes, markdown list c) Yes, Notion / other"

pool_status

pool_status

"none"

"markdown"

"notion"

"Q5: I will install a few hooks for you, reply 'yes' or press 'enter' to install:

1. Prediction lock — After we complete the prediction together, the file will be locked. Neither you nor I can modify the prediction section. Retrospective can only be appended to the lower half of the same file, without polluting the judgment in the upper half. (Without this lock, it's almost inevitable that you or I will want to 'fix the prediction after seeing the data' later — we all make this mistake)
2. SessionStart automatic report — Display buffer / pending retrospective / top candidates at the top of each new session
3. Silent usage log — Asynchronously record usage frequency, non-blocking, for future diagnosis

Install all three together. You can choose not to install (reply 'no') but you will lose the prediction lock, and the calibration value will decrease.

Reply yes / no."

hooks_installed

hooks_installed

true

"yes"

false

The most important signal source for the tool in the early stage is benchmark accounts — after you finish init without data, the equal-weight v0 rubric is like astrology. But if you can find an account you want to emulate, import 5-10 of its high / medium / low performance samples, the tool will have an anchor.

🎯 Benchmark Account

Can you find a benchmark account? At least 3 videos from this account.

  - You **have no posting history at all** (Q2=a) → **Highly recommended** — the rubric has no anchor and relies entirely on benchmarks.
    If you don't find one, use the universal v0 equal-weight rubric, and the accuracy of the first 5 pieces will be worse for longer
  - You **have posting history** (Q2=b) → **Optional** — you can also use your own history for calibration;
    But it's recommended to import at least 1 benchmark for sanity check (to see if your account truly deviates from the benchmark direction)

a) Find now → Immediately enter /cheat-learn-from (5-15 minutes, depending on how prepared your materials are)
b) Find later → Mark state as `benchmark_status: pending`, cheat-status will continuously remind you
c) Don't find → Mark state as `benchmark_status: none`, start with universal v0

Reply a / b / c.

• Select a → After scaffolding is created in Phase 3, automatically dispatch to /cheat-learn-from (don't let the user run it manually — it's already part of the init process). Return to init Phase 4 after completion
• Select b → Mark state as

  benchmark_status: pending

  +

  benchmark_name: null

• Select c → Mark state as

  benchmark_status: none

benchmark_status

benchmark_name

1. .cheat-state.json

   "Creating .cheat-state.json — the place where all sub-skills share context.
    All answers collected during this init will be written here."

   Write (all

   <...>

   placeholders must be replaced with specific enum values by checking the Q mapping table above, never store letters directly):
   json

   {
     "schema_version": "1.1",
     "skill_version": "1.0.0",
     "rubric_version": "v0",
     "content_form": "<Check Q1 mapping table, write enum string like \"opinion-video\">",
     "typical_duration_seconds": <Derived from Q1.5: 30/90/240/450/900>,
     "target_publish_cadence_days": <Derived from Q1.6: 1/2/7/null>,
     "rubric_form_mismatch": <Q1=a→false; others→true>,
     "benchmark_status": "<Derived from Phase 2.5 answer: a→\"imported\"/b→\"pending\"/c→\"none\">",
     "benchmark_name": <String name if imported, otherwise null>,
     "benchmark_sample_count": <Number if imported, otherwise 0>,
     "baseline_plays": null,
     "calibration_samples": <Q2=a→0; Q2=b→Estimated value from Q2.3 or import count>,
     "data_collection": "<Check Q3 mapping table, write \"manual\" or \"adapter\">",
     "pool_status": "<Check Q4 mapping table, write \"none\"/\"markdown\"/\"notion\">",
     "data_layer": "markdown",
     "hooks_installed": <Check Q5 mapping table, write bool true/false>,
     "enabled_trend_sources": ["manual-paste"],
     "enabled_perf_adapters": <Q2.1=a→[\"douyin-session\"]; b→[\"youtube-data-api\"]; c→[\"bilibili-stat\"]; others→[]>,
     "last_bump_at": null,
     "last_bump_self_audited": false,
     "last_published_at": null,
     "last_published_file": null,
     "last_retro_at": null,
     "last_trends_run_at": null,
     "last_trends_added_count": 0,
     "consecutive_directional_errors": [],
     "pending_retros": [],
     "shoots": [],
     "in_progress_session": null,
     "initialized_at": "<Local ISO 8601 with time zone, e.g., \"2026-05-05T20:11:13+08:00\", **do not use UTC Z suffix**>"
   }

2. rubric_notes.md

   "Creating rubric_notes.md — the real source of your scoring dimensions.
    Uses v0 placeholder rubric — equal-weight 7 dimensions (each dimension is equally important).
    
    Why call it v0: v0 is a placeholder before calibration. The true weights for your account must be inferred from your
    data, not preset. After running 5 pieces of content with data, it will automatically propose
    upgrading to \"Calibrated v1\" (your first truly calibrated rubric)."

   • Copy

     cheat-on-content/starter-rubrics/<form>-zero.md

     (cold-start) or

     <form>.md

     (still referable when there is existing data)

3. script_patterns.md

   "Creating script_patterns.md — the accumulation of your writing patterns (decoupled from rubric).
    rubric_notes.md teaches Claude how to score;
    script_patterns.md teaches Claude how to write."

   • Copy

     cheat-on-content/templates/script_patterns.template.md

4. Four directories:

   scripts/

   +

   predictions/

   +

   videos/

   +

   samples/

   (all add

   .gitkeep

   )

   "Creating four directories:
    
    scripts/      — Drafts before filming (written by cheat-seed or you)
    predictions/  — Immutable prediction logs (protected by hooks)
    videos/       — Working directory after filming (cheat-shoot creates subdirectories)
    samples/      — Benchmark account videos / transcripts (cheat-learn-from creates subdirectories)
    
    The first three use the same <date>_<id>_<short> naming convention to correlate with each other.
    samples/ is grouped by benchmark account name: samples/<account-name>/<video-id>/.

benchmark.md

"Copying benchmark.md placeholder template (actual content filled by cheat-learn-from) — 
 This is the central reference for your benchmark account.
 In the early stage, the tool's rubric / pattern / topic direction are largely derived from here;
 In the later stage when N≥10, the influence fades, but it remains for sanity check."

• Copy

  cheat-on-content/templates/benchmark.template.md

  →

  <user-repo>/benchmark.md

• Do not create if Phase 2.5 selects c → benchmark.md does not exist, mark state as

  benchmark_status: none

1. WORKFLOW.md

   +

   STATUS.md

   • Copy corresponding files from templates/
2. If Q5=yes → Install hooks
   • Read

     .claude/settings.json

     (create empty

     {}

     if it does not exist)
   • Merge

     hooks.PreToolUse

     from

     hooks/prediction-immutability.json

   • Merge

     hooks.SessionStart

     from

     hooks/session-start.json

   • Merge hooks from

     hooks/meta-logging.json

     (if enabled at the same time)
   • Copy

     prediction-immutability.sh

     +

     session-start.sh

     +

     log-event.sh

     to

     .cheat-hooks/

     , chmod +x
   • Use

     ${CLAUDE_PROJECT_DIR}/.cheat-hooks/

     as the command path in settings.json
3. (Pool option c—Notion) Only record

   pool_status: notion

   to the state file, handle it later when cheat-trends is called

1. Create video folder:

   videos/<date>_<id>_<short>/

   • <date>

     = Actual video release date

   • <id>

     = 12-digit hash, generated by sha256 for (title + platform ID)

   • <short>

     = First 3-8 characters of the title
2. Write report.md: Fill in data fetched from the adapter (views / likes / comments / shares / top comments)
3. Ask user for original script: "Do you still have the original script for video "{title}"?"
   • Yes → User provides → Save as

     videos/<id>/script.md

   • No → Mark

     script_lost

     (still create the video folder, but script.md is missing)
4. Write reconstructed prediction:

   predictions/<date>_<id>_<short>.md

   • Header marked with

     **Reconstructed retrospective — NOT a blind prediction**

   • 7-dimensional scoring is reverse-calculated based on script + actual retrospective data — clearly state this is for non-calibration purposes
   • Do not count towards calibration_samples (this is imported history, not accumulated calibration samples)

• Derive

  baseline_plays

  = Median views of fetched videos → Write to state file
• Derive confidence level → Directly use when writing predictions in subsequent cheat-predict
• Output summary: "Imported N pieces of history. The latest one has Xk views, median Yk views, created N video folders + reconstructed predictions"

1. Create temporary file

   predictions/_test_hook.md

   , containing

   ## Prediction\n[test]\n## Retrospective\n

2. Attempt to Edit the

   ## Prediction

   section of this file
3. The hook should exit 1 to block → Report "✅ Immutability hook is effective"
4. Delete the test file
5. SessionStart hook verification: Directly call

   bash .cheat-hooks/session-start.sh

   → Should output a report (even if it's empty)

1. Do not pretend success: If any step fails → Clearly tell the user which step went wrong. Never write "✅ Initialization completed" if it's not actually completed
2. Do not ask in batches: Ask the 5 questions one at a time
3. Do not silently mkdir: Explain the purpose of each file when creating it
4. Do not force SQLite: Provide all users with markdown, just mention "Will recommend upgrading when reaching 30 pieces"
5. Unified state fields: Remove enum fields such as mode / prediction_complexity / bucket_scheme — Use only calibration_samples integer + confidence derivation
6. Import failure does not block: If Q2=b but adapter installation fails / fetch fails → Gracefully degrade to "Mark calibration_samples as estimated value, do not import historical video folders"

• "Skip Q1-Q5, just create all files for me" → Refuse. The answers to the questions directly affect default configurations (content_form, cadence, hooks)
• "I've already initialized elsewhere, sync the configuration from that project" → Be cautious. Prompt the user to manually cp existing

  .cheat-state.json

  and

  rubric_notes.md

  , do not automatically sync across projects
• "Don't install hooks but keep the immutability promise" → Allow, mark state as

  hooks_installed: false

  , cheat-status will continuously prompt "Your immutability is a gentleman's agreement"

• After completion, the main SKILL.md route unlocks all other sub-skills

• cheat-status

  reads the

  calibration_samples

  field in

  .cheat-state.json

  to determine which confidence level to display
• If Q2=b and import is done → Historical reconstructed predictions are stored in

  predictions/

  and

  videos/<...>/

  , but do not count towards calibration_samples (not true calibration samples)

• /cheat-seed

  reads all historical reconstructed predictions in

  predictions/

  → Know "what the user has done before" during brainstorming