Marketing guru

marketing-guru — orchestrate the growth loop

Make a channel go viral by running a closed feedback loop backed by a persistent journal. You don't guess once — you bet, deploy, measure against the channel's own history, learn, and bet again. The creative steps are their own lego-block skills any agent can use alone; this skill is the conductor, and it also owns the thin read/pick/report helpers.

   ┌──────────────────────────────────────────────────────────────┐
   │  IDEATE  → BACKLOG → DEPLOY → (wait 48-72h) → MEASURE → LEARN  │
   │     ▲                                                      │   │
   │     └──────────── strategy + next_seeds ◀─────────────────┘   │
   └──────────────────────────────────────────────────────────────┘

The skills (invoke any one directly)

Step	Skill / here	One-liner	CLI it drives
setup	youtube-branding	channel identity: banner/avatar/logo + keywords/description (once / on rebrand)	studio brand
0	guru: journal	read phase + strategy + bets (start here)	journal · recall
1	marketing-ideate	agent generates falsifiable bets → backlog	add (· ideate fallback)
2	guru: backlog	pick the next bet (bandit, 60/40 fallback)	backlog · bandit
3	marketing-deploy	produce+publish via film-maker, then link	studio run · link
4+5	marketing-measure-learn	score virality, snapshot age buckets, slice hidden relations, then reflect → strategy (wait 48-72h)	measure · snapshots · insights · slice · compare · strategy
—	guru: report	write the growth brief to disk	report
⟳	marketing-autopilot	run the WHOLE loop on a schedule (deferred-measurement aware)	tick · autopilot

Design rule: the thinking (what to make, which bet, what was learned) is the agent's, done in the skills; the studio marketing CLI commands are helpers — pure I/O for persistence (add/strategy), retrieval (recall/backlog/journal/bandit), and deterministic work (link, measure). ideate/learn keep scripted LLM fallbacks for quick non-agent passes.

Run the whole loop

0. youtube-branding (once, before the loop) — if the channel has no banner/avatar/logo yet (or is rebranding), generate the brand kit with studio brand <spec.json>.
1. journal (below) — read the current phase + direction first.
2. marketing-ideate — web-search trends + recall winners → write bets to the backlog.
3. backlog (below) — pick the next bet.
4. marketing-deploy — produce + publish (film-maker), sized to the per-video budget, then link.
5. wait 48-72h+ for watch time to accrue.
6. marketing-measure-learn — score, then reflect into strategy.
7. Back to 2 — now exploiting what won. report (below) snapshots the cycle.

To run all of this hands-off on a schedule, use marketing-autopilot — it asks the engine what's due each tick (measure / learn / ideate / produce / idle) and does that one action, handling the 48–72h measurement-maturation wait for you.

The cold-start rule (stated once, here): relative virality is meaningless until ~10 videos exist. Deploy the first 10 as diverse EXPLORATION bets; only then does measure rank winners and learn start exploiting. The journal tracks the phase automatically (Journal.in_cold_start). Full memory + phase model: docs/50-marketing/memory.md.

Everything is per-channel — pass --channel <name> (mirrors the publish OAuth token token_<name>.json). Omit for the default journal.

journal — read the loop's state (read-only)

The first thing to run before any decision, so bets reflect the current direction.

studio marketing journal --channel <name>            # phase + strategy + every bet's outcome table
studio marketing backlog --channel <name>            # planned (queued) bets + explore/exploit mix
studio marketing recall  "<query>" --channel <name>  # past MEASURED bets most relevant to a query

You'll see the phase (COLD START n/10 vs OPTIMIZING), the Strategy (niche, current_direction, winning/losing patterns, next_seeds), and a table of every bet (id · status · idea · virality · percentile · outcome). Durable store: runs/_marketing/<name>/journal.json (machine truth) + journal.md (human render, regenerated on save — never hand-edit). Per-video snapshots: runs/<run_id>/08_stats.json, 08_comments.json.

backlog — pick what to make next

The backlog = every journal entry with status: planned. Review it, then choose the next bet to hand to marketing-deploy. The CLI lists; YOU decide.

studio marketing backlog --channel <name>   # planned bets tagged explore/exploit + counts
studio marketing bandit  --channel <name>   # the SHIPPED selector's learned theme/tag win-rates

• Primary selector: the shipped Thompson bandit (T8) over theme+tags — studio marketing bandit shows what it favors and how it ranks the backlog; the autopilot picks with it.
• Fallback heuristic (manual picks): ~60% exploitation (proven winning patterns) / ~40% exploration (fresh themes) — use this when you're choosing by hand rather than via the bandit.
• Cold-start (<10 deployed): everything is exploration regardless of tags — no baseline yet.

Sanity-check a candidate isn't a near-dupe of a past loser with recall "<idea/theme>". Empty backlog → run marketing-ideate first. No delete command yet (roadmap T2) — edit runs/_marketing/<name>/journal.json to prune, or leave stale bets unpicked.

report — write the growth brief

studio marketing report --channel <name> --provider <llm>

Writes runs/_marketing/<name>/report.md — the channel's phase, current direction, winners/losers with their assumptions (held or refuted), and the next bets. Run it after marketing-measure-learn so it reflects the latest cycle. For a quick interactive read instead of a file, just run studio marketing journal.

analysis — find hidden relations

For strategy changes that depend on production choices, do not infer from the flat journal table alone. Ask the CLI for age-normalized slices across theme, cost, music, sound, animation, effects, and providers:

studio marketing due-snapshots --channel <name>
studio marketing snapshots     --channel <name> --buckets 1,3,7,14,30
studio marketing insights      --channel <name> --json
studio marketing slice         --channel <name> --bucket 7d --group-by theme,effects,animators --metric virality
studio marketing compare       --channel <name> effects=glitch --bucket 14d --metric virality
studio marketing export        --channel <name> --format csv

Use 1d for early hook/packaging velocity, 3d for first maturation, 7d for retention and durability, 14d for stronger winners/losers, and 30d for long-tail formats. Treat every slice as association until repeated across enough examples; low n is a hypothesis generator, not a rule.

Setup check

cd /Users/dasein/dev/slope-studio
source .venv/bin/activate 2>/dev/null || { uv venv && source .venv/bin/activate && uv pip install -e ".[fal,youtube]"; }
studio marketing --help
studio yt-channel --channel <name>   # confirm WHICH channel the token points at

Always report

The channel's phase (cold-start N/10 vs optimizing), latest winners/losers with their assumptions (held or refuted), the current direction, and the next concrete bets. Be honest when a bet's assumption was refuted — that's the point.

Deeper references

• Memory & phases: docs/50-marketing/memory.md — the canonical reference for how the journal/recall/strategy/cold-start persist and self-improve.
• references/loop.md · references/scoring.md · references/analytics.md · references/trends.md.
• Architecture + full command reference: docs/50-marketing/.
• Producing videos: the film-maker skill. Why this shape: docs/20-research/self-improving-loop.md.