Weekly Learning Skill

Purpose

Make the system actually improve over time — retention, scripting,
editing, and topic choice — based on real data, not assumptions about
what "should" work better.

Process (run weekly, per channel)

1. Pull real retention data. Use the YouTube Analytics API's audience
   retention metrics (audienceWatchRatio and relative retention curves)
   for the week's videos. Identify where viewers actually drop off —
   specific timestamps/sections, not a vague "engagement was low."
2. Correlate drop-off with the actual script/edit. Check what was
   happening in the video at the drop-off point (which script section,
   which visual style moment, which SFX/VFX choice) — this ties a real
   number to a real content decision, not a guess.
3. Competitor benchmarking pass. Separately, search for what real
   top-performing channels in the same niche are doing differently this
   week (structure, pacing, hook style, topic angles) — grounded in actual
   searched examples, not general assumptions about "what works."
4. Turn findings into concrete adjustments:
   • Update script-writer guidance for that niche (e.g. shorten a
     section type that's consistently losing viewers, strengthen hooks
     that match what's retaining well)
   • Update remotion-render/vfx-audit settings if a visual choice
     correlates with drop-off
   • Feed topic-angle findings into niche-discovery/trend-research for
     future topic selection
5. Log every adjustment with the retention number or competitor
   example that justified it, per channel per week — this is the audit
   trail that proves nothing was changed on a guess.

Rules

• Never make an adjustment without a specific retention number or a
  specific real competitor example behind it — "this might work better"
  is not a valid basis for a change here.
• Don't over-correct from a single week's data on a low-sample channel —
  note sample size/confidence, and prefer patterns that hold across
  multiple weeks before making a bigger structural change.
• Keep per-channel and per-niche learnings separate — what works for one
  niche's audience may not transfer to another.