creator-teardown

Creator Teardown

Paths: ${CLAUDE_SKILL_DIR} means the folder containing this SKILL.md. On Windows, use py wherever these commands say python3.

Every question to the user goes in the question box: call the AskUserQuestion tool (2-4 options, the recommended one first). Only in an agent without that tool, ask numbered questions in text.

Output

creator-teardowns/<handle>/: teardown.md (the analysis), teardown.html (the page the user looks at), look.md (what Claude reads), style.json (every measured number, for editing in this style), plus the per-video files listed in ${CLAUDE_SKILL_DIR}/references/outputs.md. An edit plan goes to edit-plans/<name>/. Whether an edit in this style matches: scripts/match.py (references/outputs.md "Match check").

First: doctor

python3 "${CLAUDE_SKILL_DIR}/scripts/fetch.py" doctor

Run it before the first teardown in a session. Ready. means go. If it prints Not ready, or the user asks to install or set up the skill, follow ${CLAUDE_SKILL_DIR}/references/setup.md. The ElevenLabs and Gemini keys are optional. Never ask for either in the chat: the user saves them with setkey (setkey --gemini) in their own terminal.

VPY="${CLAUDE_SKILL_DIR}/scripts/run.py"; python3 "$VPY" below (py on Windows) runs a script with the tool venv's Python wherever AI_EDITOR_HOME puts it.

Pipeline

Quick mode (called by ai-editor's start, setup or style-edit)

When another skill needs only the style.json, run exactly: Step 1 list, Step 2's commands with --top 5 (5 plus 2 control and 2 spares; pasted links: all), Step 3b's four, then Step 3c's four (a gemini.py that exits 2 with no key is skipped), then stop. The Step 0 parts question only if the caller has not asked it. Skip metrics.py, the Step 3 written passes and Steps 4-6. About 3 minutes per creator on a 14-core Mac (2-3 min measured; the spares add about 1 min at worst), twice that on an 8-core laptop. Say so first.

Step 0: Ask, once

Skip what the message already answers. One question box call:

1. Which creators? The handle or links they named (Recommended) / I'll paste more. Several is

normal: people like one creator's captions and another's pace. Each creator's style is one format.

1. With more than one: which part from whom? Not sure yet (Recommended): ask again after

Step 3b, with each creator's look.md to compare; or captions, pace (cuts, zooms, motion), visuals (graphics, face framing, cards), sound.

Step 1: Get the videos

python3 "${CLAUDE_SKILL_DIR}/scripts/fetch.py" list <handle> --platform tiktok --limit 40   # or youtube
python3 "${CLAUDE_SKILL_DIR}/scripts/fetch.py" list <name> --urls picks.txt   # pasted links, one a line

Prints a ranked table (views, x median) and writes videos.json; * marks 3x the median.

Instagram: list <handle> --platform instagram reads the newest reels with views through the user's instagram.com login in their browser (sent only to instagram.com). No login: it exits 1 with one sentence; say it, or put 6+ reel links in picks.txt for --urls (ranked by likes; say so).

Step 2: Download and transcribe the winners and a control group

python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/visual.py" download <handle>
python3 "${CLAUDE_SKILL_DIR}/scripts/fetch.py" transcribe <handle>

--top N downloads and transcribes N plus 2 control videos (the 2 closest to the median; --control changes the 2) plus 2 spares for the format check (--spare), all of the creator's usual length when enough are. Pasted links: every one. The default --top 8 is 12 videos; --top 5, the fast run, is 9; pass it to both commands. Both print the count first. --ids <id>,<id> picks exact videos. The median ones are the control: a move that appears in the 500K video and the 3K video is that creator's habit, not the reason the 500K one worked. Keep at least 2.

Run transcribe and python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/gemini.py" look <handle> --no-merge at the same time (two Bash calls in one message): the model watches while Whisper listens. Every measuring script works on several videos at once (CT_JOBS=1 turns that off on a small laptop).

Say the time first (download plus transcription, about 10 s a video on a 14-core Mac). Whisper (free, local) is the default; with an ElevenLabs key transcribe uses Scribe, which keeps every filler ($0.22 an hour of audio; give the estimate before 10+ videos). If filler counts lean on Whisper transcripts, say so in the teardown.

Step 3: Measure, then analyse

python3 "${CLAUDE_SKILL_DIR}/scripts/metrics.py" <handle>

Writes metrics.json (WPM per third, pauses, fillers, sentence length, you/I/we, repeated phrases), all off the word timestamps.

Run Step 3b first, so the passes skip the videos format.json left out. Then load ${CLAUDE_SKILL_DIR}/references/analysis-framework.md and run its passes. For pass 1, launch one video-analyst agent per video, all in one message (they run at once): give each the absolute creator-teardowns/<handle>/ path and one id. Each returns a JSON beat map; merge them into the tables. Without an Agent tool, read transcripts/<id>.timed.txt yourself, one video at a time. Never read transcripts/<id>.json or metrics.json whole:

1. Beat map + timings. Every video split into labelled beats with second-marks and word counts.
2. Voice/vocab profile. Sentence length, person, commands, recurring phrases, and the moves the

user would not say.

1. Visual pass. Always: Step 3b. It is what lets the user edit their own video in this style.
2. Card events, only when the ask covers the creator's cards and graphics, or an edit plan is

coming. Writes events.json in the shape in ${CLAUDE_SKILL_DIR}/references/edit-plan.md. scripts/strip.py needs ffmpeg, numpy and pillow.

Do not skip the timings: "hook lands by 0:03 and runs 14 words, items are 8-11 s each" is a format someone can film; "hook, then list, then close" is not.

Step 3b: The visual pass (measured, no images for you to read)

python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/visual.py" measure <handle>
python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/look.py" measure <handle>
python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/format_check.py" check <handle>
python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/gemini.py" look <handle>

Run them after transcribe (look.py matches OCR text to the spoken words).

1. visual.py: cuts and their kinds, shot lengths, zooms, pans and shake.
2. look.py: OCR 3 frames a second and YuNet faces: the captions (place, size, words,

case, colours, weight, entrance), face framing, graphics share, palette, motion personality. About 40 s a video.

1. format_check.py: a style is ONE format, not a blend of all they post. It groups the videos by frame,

face, captions, shot length and length; a majority group of 3+ is kept, the rest move to video/left-out/ (format.json says why; references/look-pass.md). Tell the user its first line when any differ ("Using 4 of 6: 2 look like a different format."); if mixed, ask for more videos of one kind.

1. gemini.py look: merges what Gemini Flash-Lite said in Step 2 (it watches any video not done yet)

about the 5 most-viewed: font, graphics style, transitions, what is not generic. Never a number.

Then read `look.md` (under 48 lines). Do not open frames or sheets. Every field is in ${CLAUDE_SKILL_DIR}/references/look-pass.md. Numbers in style.json come from code; never edit them by hand. Where look.md and the model's words disagree (say motion: punchy measured, smooth named), the measured one wins and the teardown says both.

No Gemini key (gemini.py look exits 2): offer the free key once (references/setup.md step 4). If the user skips it, run the fallback: one Agent call with model: "haiku" that reads sheets/<id>.jpg for each kept video (one 3 x 3 sheet each, made by visual.py) and writes video/<id>.look-ai.json per gemini.py prompt. Exact brief in references/look-pass.md. Then python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/gemini.py" merge <handle>.

No OCR engine (doctor says optional): captions stays unmeasured and look.md says so. Install it (doctor prints the line) rather than reading sheets yourself.

Step 3c: The deep pass and the page

python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/graphics.py" measure <handle>
python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/gemini.py" kinds <handle>
python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/sound.py" measure <handle>
python3 "$VPY" "${CLAUDE_SKILL_DIR}/scripts/report.py" build <handle>

Each measure prints one summary line and the file it wrote (--json prints the full block; not needed: look.md has the numbers). Run them after Step 3b (graphics.py masks the speaker with look.py's faces). They add every graphic (cropped, its kind, where it sits, how it enters and exits as a fitted GSAP ease), cut kinds and camera moves (visual.py), sound effects and music, the first 3 seconds of each video, winners against the control group, and an AI-tell scan of the creator's own look. Fields, methods, ceilings and cost: ${CLAUDE_SKILL_DIR}/references/teardown-page.md. No Gemini key: gemini.py kinds exits 2; the kinds stay the code's guess, or run the haiku fallback in that reference.

Then open creator-teardowns/<handle>/teardown.html for the user (open on a Mac, start on Windows, xdg-open on Linux) and offer to publish it as an artifact. Do not read the page or its images yourself: look.md has the numbers. Tell the user in two or three lines what the winners do differently (the Winners: lines in look.md, with n) and any BAN-level AI tell in their look.

Then ask in the question box (the page only shows the evidence; it has nothing to tick), one multi-select question: "Which parts of @<handle> should your edit copy?" Options: captions, pace, graphic kinds, entrance motion, layout, sound (at most 4 options a question: split into two questions if needed), each option's description one measured line from look.md (e.g. "slides in 0.25 s with overshoot"). If any BAN tell sits in a part they pick, say which before they confirm. Save the answer:

python3 "${CLAUDE_SKILL_DIR}/scripts/editplan.py" blend <handle>-mine --from <handle> \
    --take captions,pace,graphics,entrances,layout,sound    # only the parts picked

style-edit reads creator-teardowns/<handle>-mine/style.json; a part left out uses its default. report.py build and editplan.py blend print the style's skill (skill edit-like-<handle> -> <path>, with ai-editor installed): tell the user its name. Feedback on that style edits that skill.

Step 3d: Several creators (when the user named more than one)

Follow ${CLAUDE_SKILL_DIR}/references/blend.md: Steps 1 to 3b per creator, look.md side by side, then editplan.py blend.

Step 4: Write the teardown

Fill ${CLAUDE_SKILL_DIR}/references/teardown-template.md into creator-teardowns/<handle>/teardown.md. Every video gets a stats line, verbatim transcript, beat table with timings, patterns to keep and patterns to reject.

Step 5 (optional): Turn it into a skill

Ask whether the format is worth keeping. If yes, follow ${CLAUDE_SKILL_DIR}/references/skill-handoff.md (the format skill, beside the style's own skill).

Step 6 (optional): Plan the user's video in the creator's style

When the user hands over their own video, follow ${CLAUDE_SKILL_DIR}/references/edit-plan.md. It needs the creator's events.json from the visual pass; it does not render.

If a script stops

Hard rules

• Real numbers or nothing. Every claim about what works cites the view count next

to it. Never say a video "went viral" without the number.

• Verbatim transcripts only. Never summarise a video that was not transcribed.
• Structure is copied, voice is not. Every teardown states what to reject.
• Median control group. At least two mid-performers transcribed, or the analysis

cannot separate the format from the creator's habits.

• Never invent a transcript line, or a stat or quote for a card. If transcription

failed, retry with --ids, then say it failed.

• Never ask for the ElevenLabs or Gemini key in the chat.
• Measured numbers only in style.json. pace and zoom come from visual.py;

captions, face, graphics and motion from look.py; only words (look, font_match) from the model.

• Read look.md, not frames. Open a frame only to settle a question look.md

cannot answer, and say which one. The user gets teardown.html; Claude gets look.md.

• Winners against control: say n. Never call a difference proven; "every winner"

only means no overlap in this sample.