analytics-feedback

name	analytics-feedback
description	Close the learning loop from real YouTube performance. Reads the newest YouTube Studio "Table data.csv" export, attributes each video's views/CTR/retention to topic + named-entity tokens, and rewrites the AUTO-managed block of schedule-drip's topics.scorelist (GO winners / HOLD dead niches) plus appends proven-winner search queries to scout-sources' niches.txt. Deterministic, no Claude, idempotent (no-op unless the export's mtime changed). The analytics half of the autopilot loop — runs at the top of every autopilot tick so scout + schedule-drip always reflect what actually performed.
allowed-tools	Bash
user-invocable	true

analytics-feedback

Turns channel analytics into pipeline behavior. YouTube Studio → Analytics → Content → Export → unzips to ~/Downloads/Content <range> <channel>/ with Table data.csv (per-video views, watch hours, CTR). This skill reads that export and re-derives which topics to chase and which to bury.

Usage

analytics-feedback.sh                       # newest export under ~/Downloads/Content*C0BALT_CUT*/
analytics-feedback.sh "/path/Table data.csv"   # explicit file
analytics-feedback.sh --force               # re-learn even if the export's mtime is unchanged

Idempotent: records the consumed export's mtime in work/_autopilot/analytics.csv.mtime and no-ops on an unchanged export (so it's safe to call every autopilot tick). Drop a fresh export into ~/Downloads and the next tick relearns.

What it writes

1. .claude/skills/schedule-drip/topics.scorelist — preserves everything ABOVE the # ==== AUTO (analytics-feedback) ... sentinel (your hand-curated rules are never touched) and regenerates the block below it with data-derived GO <pattern> / HOLD <pattern> lines. Evidence (n=, med=, ctr=) is on # comment lines — never inline on a rule line, because schedule.py treats everything after the verdict as the regex.
2. .claude/skills/scout-sources/niches.txt — preserves your manual seed queries and appends a regenerated AUTO block of search queries built from the GO winners (e.g. a winning black[ -]?hole → black hole physics explained), so scout expands into proven niches. Additive only.
3. work/_autopilot/topic_scores.json — full per-token evidence (n, median/mean/min/max views, mean CTR, median retention, total watch hours, verdict, suppressed-alias flag, manual conflicts). The audit trail.

Classification (env-tunable)

Per token (known entities from a lexicon + auto-discovered title tokens, deduped so #blackhole doesn't double-count black[ -]?hole):

• GO when n ≥ AF_GO_MIN_N (3) AND median_views ≥ AF_GO_VIEWS (600) AND mean_ctr ≥ AF_GO_CTR (5%). Median + the n≥3 floor reject a single viral fluke (a 2-video token with one 6-view dud can't earn GO).
• HOLD when n ≥ AF_HOLD_MIN_N (2) AND median_views ≤ AF_HOLD_VIEWS (60). Dead-is-dead needs less evidence than proven-winner.
• neutral otherwise (no line emitted; schedule.py defaults unmatched sources to HOLD anyway).

A HOLD always wins over a GO in schedule.py, so the data can demote a manual GO (surfaced as a conflict in the JSON) but a manual HOLD veto still sticks.

Where it runs

• autopilot tick — first step of autopilot.sh, before scout, so discovery + staging always reflect the latest export.
• standalone — run it by hand after any fresh export to retune immediately.