Shellwire.

The problem on file

Shellwire ingests posts from a social network, enriches them with editorial AI, scores emergence across the corpus, and republishes the result as a newspaper-style web product with a newsletter, comments, programmatic SEO, and a daily cadence.

It must do this 24 hours a day, with minimal human intervention, on a near-zero AI budget.

The hard part is not "call an LLM." The hard part is keeping that pipeline alive when one provider rate-limits you, one returns chain-of-thought leakage, one is banned for a billing problem, and one is simply slow. The interesting work is the orchestration around the model, not the model.

What it is, in numbers

Exhibit A·Captured 2026-05-21 07:38 UTC·Source shellwire.comlive

The marquee piece, the council

Shellwire's editorial AI does not call one model. It calls a council of twelve free-tier providers, picks a strategy per task, and records every vote.

The orchestrator routes by task, not by model: classification jobs use majority-vote, the weekly narrative uses synthesize-best, the GSC growth report uses ranked-fallback at council size ten because it has to survive provider rate limits.

The strategy switch is the point of the system. Four lines of routing logic that turn twelve flaky free-tier endpoints into one reliable editorial brain.

switch (strategy) {
case 'single-query':
  result = await this.singleQuery(request, providers);
  break;
case 'majority-vote':
  result = await this.majorityVote(request, providers);
  break;
case 'synthesize-best':
  result = await this.synthesizeBest(request, providers);
  break;
case 'ranked-fallback':
  result = await this.rankedFallback(request, providers);
  break;
default:
  result = await this.singleQuery(request, providers);
}

result.totalLatencyMs = Date.now() - start;
this.analytics?.logQuery(request.task, request.tier, request.prompt, result);

The hidden ninety percent

Ingestion, end to end

The pipeline runs every four hours. It expects every dependency to fail at least sometimes, and is structured around that expectation.

The three-tier source fallback (Moltbook API, then markdown discovery files, then Brave Search) is the one I am proudest of. The first time it survived a Moltbook outage in production with no intervention, that was the morning I knew the architecture was correct.

Heat, quality, and the daily hero

Stories are scored on a heat formula tuned for editorial behaviour: likes + 3 × comments + keyword_bonus, with a 150-point bonus for world events and a 50-point bonus for AI safety topics. The AI editorial pass attaches a quality score that boosts heat by up to 20 percent.

The Daily Hero is selected transactionally with a seven-day fallback window so the front page never goes blank.

Distribution as part of the product

The growth-worker is a separate service. It listens for ingestion:complete on Redis pub/sub, regenerates the sitemap, validates its integrity (no orphan slugs, no error-AI articles, no junk), and pings IndexNow. Programmatic SEO is not an afterthought, it is wired into the publishing loop.

In production