guerrilla-marketing

guerrilla-marketing — indirect reach via grounded comments

studio guerrilla finds recent uploads on a watchlist of other creators' YouTube channels, writes one comment that reacts to something specific the video actually says, scores it with a critic, runs it through a set of rate-discipline rails, and — only if everything passes — posts it from the configured channel. It then tracks whether the comment survives and, over weeks, whether any of this moves subscribers.

The mechanism is indirect. Nobody clicks a link — there isn't one; the deny-list makes posting one a code-level impossibility. The bet is that an interesting, on-topic comment makes a reader click the commenter's name, which lands them on the configured channel's own page. That's the whole conversion path: one good comment, one profile click, maybe one subscribe.

This is gray-hat and ToS-adjacent. Automated commenting is spam under YouTube's Terms of Service under any honest reading of the policy — nothing here asks permission or hides behind a gray area. It just tries to stay far enough inside YouTube's actual enforcement behavior (comment removal, shadow-holds, account flags) to not get the channel banned. Every comment posts through the operator's own OAuth client, so every comment is attributable to the real channel — there is no throwaway identity absorbing the risk. Ban-defense is therefore the core design, not an afterthought: a hard daily cap, per-channel cooldowns, a denylist that makes self-promotion impossible in code (not just an LLM instruction), a near-duplicate filter, style/opening diversity checks, and a circuit breaker that watches comment survival and pages the operator when it drops. None of that makes this compliant — it makes it a calculated, monitored bet.

The pipeline

 watchlist channels
        │
        ▼
   DISCOVER  ── new uploads per watched channel (YouTube Data API, cheap quota)
        │
        ▼
     RANK    ── hard gates: age, comment count, channel size, cooldown, daily cap,
        │        active window (cheapest gate first — nothing paid runs yet)
        ▼
  TRANSCRIPT  ── cached fetch (network, not LLM); no captions → skip at zero LLM cost
        │
        ▼
    TOPIC     ── classify against a transcript excerpt; unclear topic → skip
        │
        ▼
  HIGHLIGHT   ── find one real, specific moment in the video to react to
        │
        ▼
   COMPOSE    ── several comment variants, each a different style + wording
        │
        ▼
    CRITIC    ── grounded judge scores each variant against the transcript excerpt
        │
        ▼
    RAILS     ── denylist, near-duplicate, opening-repeat, style-overuse,
        │        timestamp-verify, spacing, active-window, blackout
        ▼
     POST     ── the OAuth client posts the #1 surviving variant
        │
        ▼
    TRACK     ── re-check likes/replies/survival; the circuit breaker lives here