| name | content-performance-audit |
| description | Statistically audit what actually drives a creator's views. Use when a SocialGPT user asks "what drives my views", "why do some of my posts pop", "what should I double down on", or wants a data-backed performance review of their own content. Pulls the user's videos via the SocialGPT MCP, then runs real significance tests (Pearson, Kruskal-Wallis) to rank the factors — length, timing, platform, format — that predict view count, and writes a visual report. Requires the SocialGPT MCP server to be connected. |
| license | MIT |
Content Performance Audit
Turn a creator's own post history into a ranked, statistically-tested answer to
"what actually drives my views?" — instead of guessing, you measure.
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: "what drives my views?", "what's working on my
account?", "what should I post more of?", "do a performance audit", "why
did some videos pop and others flop?"
Workflow
-
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.
-
Pull the data. Call the MCP tool:
list_videos(limit=50, sort="recent", include_analysis=false)
To audit a single connected account, first call list_accounts() and pass
its account_id. Save the raw JSON the tool returns to a file named
videos.json in the working directory (the whole {"videos": [...]}
envelope is fine — the script handles it).
-
Run the analysis (it has no third-party dependencies):
python scripts/analyze.py videos.json
The script reads videos.json, runs the tests, prints a Markdown summary to
stdout, and writes content-performance-report.html.
-
Relay the result. Present the Markdown summary the script printed, then
offer the user the generated content-performance-report.html to download /
open (on Claude.ai it appears as a downloadable file).
What the script does
For every video with a usable view count (Instagram posts showing 0 plays are
treated as missing, since IG hides reel plays), it tests each factor against
view count and ranks them by effect size and p-value:
- Continuous factors (video length, engagement rate) → Pearson correlation
on log-views, with a real two-sided p-value.
- Categorical factors (platform, day of week, time of day, sequel vs.
standalone) → Kruskal-Wallis rank test, robust to the heavy-tailed view
distribution.
It flags factors at p<0.05 (significant) and p<0.01 (highly significant), then
translates the top significant factors into plain-English actions. See
references/methodology.md for the statistical details and caveats.
Notes
- Honest about samples. With fewer than ~6 usable videos it will tell you so
rather than invent findings. More posts → more statistical power; 30+ is ideal.
- Correlation ≠ causation. Significant factors are the best hypotheses to
test next, not guarantees. The report says this explicitly.
- Re-runnable. Suggest the user re-run monthly; drivers shift as the account
grows.
- Part of a larger loop. This is the "find your drivers" stage of the
going-viral strategy loop — run it there to turn the
drivers it surfaces into your next experiment.