/analyze-reel

One command, IG URL in → full teardown out. Apify pulls the post, ffmpeg pulls frames, whisper.cpp transcribes the audio, the model reads it all and writes a structured breakdown to disk. No FAL, no generation — just analysis.

Flow

1. Parse URL — extract shortcode from instagram.com/reel/<code> or instagram.com/p/<code>. Derive slug ig-<shortcode>.
2. Set output dir — ~/Desktop/STUDIO/post-analysis/ig-<shortcode>/. Create it. If it already exists with a populated analysis.md, ask Will whether to re-run or skip — never silently overwrite.
3. Scrape post via Apify — use mcp__apify__search-actors with query "instagram post detail" if the actor isn't already known. Prefer apify/instagram-scraper (handles single-URL detail scrapes). Call mcp__apify__fetch-actor-details to confirm input schema, then mcp__apify__call-actor with { "directUrls": [<url>], "resultsType": "details", "addParentData": false }. Wait for the result (sync). Pull: caption, hashtags, ownerUsername, ownerFullName, likesCount, commentsCount, videoViewCount, videoPlayCount, timestamp, videoUrl, displayUrl, productType, dimensions.
4. Save metadata — write the full scrape JSON to metadata.json in the output dir.
5. Download video — scripts/download_video.py <videoUrl> <out-dir>/video.mp4. Skip if file already present and >100KB.
6. Extract frames — scripts/extract_frames.py <out-dir>/video.mp4 <out-dir>/_frames/. Outputs cover (0s), hook (1s), every 2s thereafter, plus the last frame. Also saves cover.png (frame at 0s) at the output root for convenience.
7. Transcribe audio — reuse ~/Desktop/STUDIO/3-studio-master/skills/post-content/scripts/transcribe.py <out-dir>/video.mp4. Caches as video.mp4.transcript.txt alongside the video. Copy/symlink contents to transcript.txt in the output dir for cleanliness.
8. Read everything back into context — read metadata.json, transcript.txt, and view the frames from _frames/ directly (multimodal). Cap frames read at 12 — if more than 12 exist, sample evenly across duration so the first, last, and pacing transitions are covered.
9. Write `analysis.md` — see structure below. Save to <out-dir>/analysis.md.
10. Print summary to chat — 6-8 lines: hook verdict, structure verdict, top replicable pattern, link to the saved folder. No emojis.

Analysis structure — write this exact skeleton

# IG Reel Teardown — @<owner> · <shortcode>

**Posted:** <date>
**Views:** <n> · **Likes:** <n> · **Comments:** <n>
**Duration:** <seconds>s
**URL:** <original-url>

## 1. Hook (0-3s)
What the viewer sees and hears in the first 3 seconds. Quote the opening line of the transcript. Describe the cover frame. Why it stops the scroll (or doesn't).

## 2. Structure & pacing
Beat-by-beat. Use transcript timestamps if available, else describe the frame transitions. Note cut frequency, energy curve, where attention is reinforced or risked.

## 3. Visual language
Aesthetic, color palette, framing (talking-head / b-roll / text-on-screen / split-screen / etc.), typography if any, motion type. One short paragraph.

## 4. Audio & voice
Voiceover delivery, music presence, cuts in audio, sound design. If music is dominant, name the energy/genre. If voiceover, characterise tone (urgent / dry / conspiratorial / warm).

## 5. Caption strategy
Full caption verbatim in a quote block. Then: what role the caption plays (extends the hook / context-fills / CTA / standalone story). Hashtag count and category mix.

## 6. What's working
3-5 bullets. Specific, not generic. "Cuts every 1.2s on hard consonants" beats "good pacing."

## 7. Replicable patterns
3-5 bullets. Things Will could lift into his own content. Format: **<pattern name>** — <one-line how-to>.

## 8. Verdict (creator POV)
2-3 sentences. Would Will copy this? Where does it lose him? What's the single thing worth stealing?

URL parsing rules

• instagram.com/reel/<CODE>/ → shortcode = <CODE>
• instagram.com/p/<CODE>/ → shortcode = <CODE> (treat carousels/single-image posts the same — analysis still useful for caption + cover)
• Strip query strings (?igsh=...) before extracting.
• If the URL doesn't match, ask Will to confirm the URL.

Apify actor notes

• Default actor: apify/instagram-scraper (or whichever surfaces with the best rating on search-actors).
• Input shape varies by actor version — always call fetch-actor-details first to read the current input schema. Don't guess field names.
• Cost: a single-URL details run is cheap (~$0.001). Always run sync (async: false).
• If the scrape returns empty (private account, deleted post, geo-blocked), surface the error to Will and stop — don't proceed with empty metadata.

Frame extraction rules

• Output PNG (not JPG) — lossless, easier to read fine text overlays.
• Width: 720px (downscale longer side). Keeps files small without losing legibility.
• Naming: zero-padded index + timestamp suffix → 01_t0.0s.png, 02_t1.0s.png, 03_t3.0s.png, …
• Always include t=0 (cover) and t=duration (last frame).

What this skill does NOT do

• Does not post, repost, save to camera roll, or call any Meta API.
• Does not analyze TikTok, YouTube, X, or any other platform. (Future skill.)
• Does not generate new visuals — for that, hand off to /static-ads or /thumbnail.
• Does not auto-add the post to any tracking sheet or vault. If Will wants it logged, he'll say so — then write a note under ~/Obsidian/Inbox/parked/ or the relevant project.

Failure modes to handle

• Apify actor fails / times out → retry once. If still fails, surface the actor's error and stop.
• Video URL expired (IG signed URLs decay) → re-run the Apify scrape, the URL is regenerated each call.
• whisper-cli not installed → tell Will brew install whisper-cpp and stop. Don't fall back to a stub.
• ffmpeg not installed → same — tell Will brew install ffmpeg and stop.
• No audio track (rare) → write transcript.txt with [no audio track] and skip section 4 in the analysis.