sk heavy industries

ask-the-lyrics agent active

agentsautomationcontent-pipelinesocial

One of the agents in the fleet: a content pipeline that turns the ask the lyrics engine into a self-publishing feed. It runs every few hours as a quality-gated, self-throttling loop.

 every few hours  (ask-the-lyrics agent)
        │
        ▼
 source    : Reddit + X, ~15 questions from each
        │
        ▼
 dedupe    : normalize, check against a seen-questions store
        │
        ▼
 filter    : keep only DB-answerable & share-worthy questions
        │
        ▼
 answer    : call the ask-the-lyrics API
        │
        ▼
 score     : rate each answer 1–10, keep only ≥ 7
        │
        ▼
 publish   : Q&A threads (link back), hard cap 2 pairs / 4 posts per run

The pipeline

  1. Source. Searches hip-hop subreddits and X for recent lyrics questions (~15 from each).
  2. Deduplicate. Normalizes each question (lowercase, trim, drop the trailing ?) and checks it against a seen_questions store, so nothing gets answered or posted twice.
  3. Filter for suitability. Keeps only questions a lyrics database can actually answer: specific, about songs/artists/lyrics, and interesting enough to share. It discards opinion, production/beats, fashion, and "what is rap?"-level prompts.
  4. Answer. Calls the ask the lyrics API for each surviving question.
  5. Quality gate. Scores every answer 1–10 (are specific lyrics cited? multiple artists? surprising?) and only publishes those scoring ≥ 7.
  6. Publish (rate-limited). Posts the best as question → answer threads that link back to the full answer page, with hard caps per run (at most 2 Q&A pairs / 4 posts) so it stays a trickle, not a firehose.

The interesting part isn't the posting. It's how much the agent throws away. Dedup, a suitability filter, and a numeric quality threshold mean it self-censors low-value output instead of publishing everything it can generate.

Stack: An OpenClaw agent, web + X search, the ask-the-lyrics API, cron.