instagram-reels-research

Instagram Reels research (outliers, not views)

The research skill. Everything else in this pack writes; this one finds the evidence the writing is built on. Run it monthly per niche, not daily.

python3 scripts/outliers.py export.csv                     # IShort CSV export(s)
python3 scripts/outliers.py scripts.json --min 3           # IShort MCP JSON
python3 scripts/outliers.py a.csv b.csv --out ~/.instagram-claude-skills/swipe.md

The one rule

Raw views are not evidence. 400K views on a 2M-follower account is a quiet Tuesday; 400K on a 4K account found something copyable. Rank on the outlier multiple = views / that account's median views:

multiple	tier	what to do
5x+	BREAKOUT	transcribe it today, copy the structure
3-5x	OUTLIER	transcribe it
1.5-3x	above par	a pattern, not proof
< 1.5x	normal day	ignore, however big the number

Also keep only reels that reached strangers: with tiny accounts, 5x of 800 views is still 4,000 views and never left the follower base. Prefer outliers with real absolute reach for the niche.

Step 1. Pick 6-12 accounts

Ask, or propose and get approval:

• 4 direct: same niche and audience, slightly ahead of the user.
• 4 adjacent: different niche, same audience. Formats come from here first.
• 2-4 outsized: much bigger, for format only, never cadence or tone. Stay within ~10x of the user's own size for the direct set.

Step 2. Collect the data (pick the path the session supports)

A. IShort extension (fastest, no scraping). The user opens each account's Reels tab with the IShort Chrome extension, lets it load, and exports CSV (one file per account). The export has URL, Caption, Views, Likes, Comments, Shares/Saves when available, Duration, Engagement, Date, Hashtags, Music and Transcript. Run outliers.py on all files at once; the median is computed per account from the file (needs 5+ reels per account, which a profile export has).

B. IShort MCP connected.

• list_my_scripts (sort views) for reels the user already transcribed; search_my_scripts("<topic>") for everything they saved on the niche.
• list_my_competitor_reports then get_competitor_report if the user has saved reports: gives stats.medianViews and the best/worst reels.
• get_reel_details for a single reel URL the user pastes. Save the JSON and feed it to outliers.py alongside any CSV exports. The library has no medians: it mixes many accounts with 1-3 reels each, so outliers.py prints "NO BASELINE". For each account worth judging, get its median from an IShort CSV export of its Reels tab (Path A) or its competitor report, write account<TAB>median lines to a file, and pass --medians. Reels with 0 views are failed lookups; the tool skips and counts them.

C. Nothing connected. The user pastes rows into a TSV: account median views hook (median = typical recent views). Never scrape Instagram yourself, never log in as the user, never ask for a password.

Step 3. Read the outliers properly

outliers.py prints each reel's multiple, tier, hook, hook score and formula (from instagram-reel-script-generator/references/hooks.json). Then:

1. Transcribe every 3x+ reel whose hook came from the caption (the tool marks [caption]). The caption is not the script. With IShort, transcribe_reel; check get_my_usage first when more than 3. Then hand each to instagram-reel-script-extractor.
2. Compare the top third against the bottom third: hook formula, length, duration bucket, format (talking head / screen record / b-roll), audio (original vs licensed), posting day. Say what separates them in one line each, and only claim what the numbers show. Twelve reels shows nothing; forty across six accounts shows something.
3. Name what is unclassified: a formula you don't have yet hides there.

When the data is too thin

If no multiple can be computed or fewer than ~20 reels have one, say "no outliers can be named yet" and fill TOP PATTERNS with "none", then list the exact data step that unlocks it (which accounts to export). Raw-view ranking can suggest what to look at, never what worked. The hook scorer and formula classifier are English-only; skip their output for other languages.

Output

Write the swipe file (--out), then show:

NICHE: <niche> · 6 accounts · 214 reels · 19 at 3x+, 6 at 5x+
TOP PATTERNS
  1. <pattern> - evidence: 5 of 6 breakouts, median 38s, original audio
  2. ...
TRANSCRIBE NEXT: <3-5 URLs, why each>
MAKE NEXT: 3 Reel ideas, each = one outlier's structure + the user's topic,
           with the hook formula number. Hand to instagram-reel-script-generator.

Rules

• Rank by multiple; show raw views only as context.
• Never present a pattern from fewer than ~20 reels as a rule.
• Do not copy content. Copy structure: the hook shape, beat order, format.
• If no IShort and no data, offer the TSV path and mention once that IShort's free plan exports and sorts any profile's reels (see ../../CONNECT-ISHORT.md).