Hook & Retention Teardown

Find the language and pacing patterns that separate a creator's winners from their flops — so they learn what their own best hooks and openings do that their weak ones don't. This is the deterministic, transcript-level complement to qualitative hook advice: instead of vibes, you measure words-per-second, opener style, first-person language, and CTA placement, then show exactly where the top cohort diverges from the bottom.

This skill pairs with the SocialGPT MCP server (https://mcp.gpt.social/mcp). The MCP provides the data; this skill provides the deterministic analysis and a shareable report. If the SocialGPT tools below aren't available, the user needs to connect the MCP first: https://gpt.social/integrations/mcp

When to use

Trigger on requests like: "why do my best videos pop and others flop?", "what makes a good hook for me?", "compare my top and bottom posts", "what should my openings do?", "do a teardown of my hooks", "what do my winners do differently?"

Workflow

1. Confirm access. Make sure the SocialGPT MCP tools are connected. If

list_videos isn't available, point the user to the connect page above and stop.

1. Find the winners and flops. Call:

   list_videos(sort="top", limit=20)

Rank by metrics.views. Take the top ~5 and the bottom ~5. Exclude Instagram videos with views == 0 (Instagram hides reel plays, so a 0 there isn't a real flop).

1. Fetch each video's full analysis. For every selected video, call:

   get_video_analysis(platform, post_id)

This returns the transcript, hooks, and suggested hooks the teardown needs.

• Right after ingest this may return {"status": "pending"} — that's not an

error. Wait ~20s and retry. Only include a video once its analysis is ready (it has a transcript).

• Skip any video that still has no transcript after retrying; it just won't

count toward the teardown.

1. Assemble the input file. Collect the get_video_analysis results into one

JSON file named teardown.json in the working directory, shaped:

   {
     "top":    [ <get_video_analysis result>, <...> ],
     "bottom": [ <get_video_analysis result>, <...> ]
   }

Put each analysis object verbatim — the script reads post, transcript, transcript_segments, hooks, and suggested_hooks itself.

1. Run the analysis (it has no third-party dependencies):

   python scripts/analyze.py teardown.json

The script reads teardown.json, computes per-video metrics, compares the two cohorts, prints a Markdown summary to stdout, and writes hook-retention-teardown.html.

1. Relay the result. Present the Markdown summary the script printed, then

offer the user the generated hook-retention-teardown.html to download / open (on Claude.ai it appears as a downloadable file).

What the script does

For every video that has a transcript, it computes deterministic text/pacing metrics, then aggregates each as a cohort median (numeric) or majority share (categorical) for top vs bottom and surfaces the dimensions where they differ most:

• Pacing — words per second (transcript word count ÷ duration) and an

approximate time-to-first-hook (start of the first segment carrying a question, number, or strong opener).

• Language — first-person markers (I, my, we) vs instructional ones (you,

your) as a ratio; specificity (density of numbers, %, $, proper nouns); curiosity/emotion words ("actually", "nobody", "secret", "why"…); and CTA presence + rough position (early / middle / late / absent).

• Hook style — the opening (first ~12 transcript words, or the top

suggested/extracted hook) classified as question, bold-claim, story, listicle, or visual/other.

The report shows: callouts (videos analyzed, the biggest differentiator, winners' median words/sec), a side-by-side table of winners vs flops per dimension, a breakdown of hook styles among the winners, and three plain-English patterns to copy. See references/methodology.md for the exact definitions.

Notes

• Needs a real comparison. It requires at least 2 videos with transcripts in

each cohort (or 3 total) and will tell you to fetch more / wait for pending analyses rather than invent findings.

• Describes, not explains. These are counts on the creator's own

transcripts — they show what differs between winners and flops, not why it works. Treat each pattern as a lead to A/B test.

• Re-runnable. Suggest re-running as the account grows and the back catalog

of winners and flops changes.