feed-riff

name	feed-riff
description	Generate Rick-voice opinion drafts (blog + social) from a single trending RSS feed item. Always sources up-front, opinion-not-recap, founder-direct dry humor.
metadata	{"clawdbot":{"emoji":"📡","tier":"3.6","handler":"feed_riff"}}

feed-riff

Turns one trending feed item into a publishable opinion-piece bundle in Rick's voice. Triggered by scripts/feed-poll.py when sm_algo_score > 0.9 for a shir-man source item, dispatched as a feed_riff workflow.

Inputs

A single feed_item dict (provided via workflow.context_json["trigger_payload"]["feed_item"]):

{
  "id": int,
  "source": str,         # e.g. "shir-man"
  "source_kind": str,    # e.g. "hn-best", "github-trending"
  "title": str,
  "url": str,
  "summary": str,        # 1-3 paragraphs from the source
  "sm_algo_score": float,
  "published_at": str    # ISO8601
}

Outputs (4 artifacts per riff)

1. Short blog draft (~400 words) — written to:

   • ~/meetrick-site/blog/drafts/YYYY-MM-DD-<slug>.html
   • ~/meetrick-site/blog/drafts/YYYY-MM-DD-<slug>.md
   • Manifest entry appended to ~/meetrick-site/blog/drafts/manifest.json

2. Three social variants (delivered inline in the artifact metadata, NOT auto-posted):

   • LinkedIn long-form (~1200 chars, 3-5 short paragraphs, opens with the source link)
   • Bluesky short (one punchy line + source URL, <300 chars)
   • Mastodon medium (~500 chars, 2-3 short paragraphs, opens with the source attribution)

Voice rules (non-negotiable)

• Source upfront. First sentence cites the source by name + linked title. Never bury the lede.
• Opinion, not recap. Don't summarize the article — react to it. What's the one thing Rick would say if asked at a dinner table?
• Founder-direct. First-person from Rick (the autonomous agent). No "AI says…" — "I think…"
• Dry humor allowed, never forced. A wry observation > a punchline. If the joke needs setup, kill it.
• Short paragraphs. 1-3 sentences each. White space is a feature.
• No filler openings. Banned: "In today's fast-moving AI landscape", "It's no secret that", "As we all know".
• Stake a claim. Every riff ends on a falsifiable opinion or a specific prediction. No "time will tell."

Safety / approval gates

• RICK_FEED_RIFF_AUTOSEND=0 (default) — drafts only. Vlad reviews via TIER-3.5 #12 Telegram /inbox.
• Set to 1 only after 3 successful manual reviews (per master plan ship-order step 3).
• Drafts NEVER auto-publish to meetrick.ai/blog/ — they live in /blog/drafts/ until promoted.

Handler

runtime/skill_handlers/feed_riff_handler.py — calls runtime.llm.generate_text("writing", ...) with the prompt template, parses output, writes the 3 files + appends manifest, returns a StepOutcome with all artifacts.

Cost model

~1 LLM call per riff at writing-route rates. Expected ~$0.05-0.15/riff (Sonnet). Daily ceiling depends on how many items cross the 0.9 threshold — typically 3-8 riffs/day → ~$0.50-1.50/day at steady state.