Audit_User_Agent — the brain of the Bizzal content machine

Three agents that run daily on GitHub Actions across every Bizzal channel that touches a viewer (YouTube, Instagram, and the three blogs + their X accounts).

Advisory only — none of them ever edit or publish your content. They read, measure, and critique; you (or a PR) act on it.

Time (UTC)	Agent	Arc	What it answers
11:00	scout	source	What should we make next?
12:00	measure	measure	What did viewers actually do?
13:00	reviewer	review	Was what we published any good?

They're deliberately ordered: scout gathers opportunities, measure pulls real numbers, then the reviewer critiques. Everything degrades gracefully — a dead feed, a missing key, or one bad property is reported, never fatal.

scout/     feeds.yml → fetch_feeds → rank (Claude) → opportunities/*.json + digest issue
measure/   collect_youtube (Data API) + collect_instagram (published file) → metrics/*.json + digest
reviewer/  targets.yml → collect → judge (Claude + rubrics) → per-repo issues + digest

---

Content reviewer — the review arc

Scores every property against a per-channel rubric and files the critique as GitHub issues. Targets live in reviewer/targets.yml.

Property	Live source	Repo source	Rubric
Bizzal Games — YouTube	RSS feed for @Bizzal_Games	— (Supabase-driven)	youtube
Bizzal Games — Instagram	published insights (see below)	—	instagram
It's Already Written	blog fetch · X deferred	_posts	blog
It's Already Priced	blog fetch · X deferred	_posts / _field_notes	blog
It's Already When	blog fetch · X deferred	_posts / _field_notes	blog

deferred = no free way to read it (login/bot walls). The property is still reviewed on its other sources, and the judge is told not to score criteria it couldn't actually see — an unavailable source never drags the score down.

Instagram is fully reviewed. IG can't be read for free, and personal-account secrets don't cross repos — so Bizzal-Games-YT-PUB (where the IG token already lives and posts) collects insights into its data/metrics/instagram.json, and reviewer/instagram.py reads that published file. No IG secret is needed in this repo.

How it reports

• a detail issue per property in its own repo (via BIZZAL_REVIEW_PAT), so findings sit next to the code that produced them — one stable issue, appended daily;
• a digest issue here, authored by github-actions[bot] via the default GITHUB_TOKEN — that authorship is deliberate: GitHub never emails you about your own PAT-authored actions, so the bot-authored digest is the one that actually reaches your inbox. It links out to every detail issue.

Rubrics live in reviewer/rubrics/: youtube · instagram · blog · blog_growth

blog_growth is an alternate lens — it ignores editorial quality and scores purely on clicks (CTR) and follower conversion. Apply any rubric to any run with the rubric input (see below).

---

Viewer metrics — the measure arc

Pulls real numbers so review stops being opinion and starts being evidence. Writes a time-series to measure/metrics/<channel>.json (committed daily) and files a metrics digest issue.

Channel	Source	Metrics
YouTube	Data API v3 (uploads playlist → videos)	views, likes, comments
Instagram	data/metrics/instagram.json published by Bizzal-Games-YT-PUB	reach, views, likes, comments, saves

Video IDs come from the uploads playlist via the Data API, not the public RSS feed — the feed 404s intermittently even from Actions.

Not yet wired: YouTube retention/CTR/watch-time (needs the Analytics API + OAuth) and blog analytics (GitHub Pages has none; GoatCounter is the plan).

---

Content scout — the source arc

Unifies content discovery. Where the reviewer critiques what's published, the scout finds what to make next.

• scout/feeds.yml — one registry of every RSS/Atom/social feed, tagged by topic (ttrpg / crypto / cyber). A feed shared across beats (e.g. The Hacker News → crypto + cyber) is fetched once and routed to every channel that wants it.
• Per-channel opportunities — ranked by Claude into concrete title/hook ideas with a why-now and a source link.
• Cross-channel plays — one story that feeds 2+ channels (a supply-chain hack for crypto-security and cyber), with the per-channel spin. This is the point of unifying: the four pipelines were previously blind to each other.

Output: machine-readable queues at scout/opportunities/<channel>.json (for a pipeline's create step to consume) plus a daily digest issue. Advisory today — not yet wired into the pipelines' generation step.

---

Secrets / variables

Repo → Settings → Secrets and variables → Actions. bizzal70 is a personal (non-org) account, so secrets do not carry over from your other repos — each must be added here, even if the same value already exists elsewhere.

Secrets

Secret	Used by	Required?
ANTHROPIC_API_KEY	reviewer + scout (the judge/ranker)	yes
BIZZAL_REVIEW_PAT	reviewer — fine-grained PAT, issues:write on all five repos	yes for per-repo detail issues; without it you still get the digest
BIZZAL_YT_DATA_API_KEY	measure — YouTube Data API v3 read key (AIza…)	yes for YouTube metrics

No Instagram or scraper key is needed here — IG metrics arrive via the games repo's published file, and all other collection is key-free.

Variables (optional)

Variable	Default	Notes
REVIEW_MODEL	claude-sonnet-5	set claude-opus-5 for the sharpest critique
SCOUT_LOOKBACK_DAYS	3	how far back the scout reads feeds
YT_CHANNEL_ID	UCn8fIswollQTSAJYkAshjyw	@Bizzal_Games

Unset Actions variables arrive as an empty string, not as absent — the code coalesces blanks to the defaults above with or, not dict.get(k, default).

---

Run it

Workflow	Schedule	Manual inputs
daily-content-scout	11:00 UTC	—
daily-viewer-metrics	12:00 UTC	—
daily-content-review	13:00 UTC	only, rubric

Reviewer inputs

• only — one property id or a comma-separated list; blank = all. Ids: bizzal-games-youtube, bizzal-games-instagram, itsalreadywritten, itsalreadypriced, itsalreadywhen
• rubric — override the rubric for every reviewed property; blank = each property's default. e.g. blog_growth for a clicks/followers pass.

# score just the crypto blog on the growth lens
gh api -X POST repos/bizzal70/Audit_User_Agent/actions/workflows/daily-review.yml/dispatches \
  -f ref=main -f "inputs[only]=itsalreadypriced" -f "inputs[rubric]=blog_growth"

Local debug

REVIEW_ONLY=itsalreadypriced python -m reviewer.run
python -m scout.run
python -m measure.run