Scribe

What Scribe does

Scribe owns the learn library — the guides on AI agents, automation, and trading basics that make up most of what search engines and LLMs cite this site for.

The loop runs nightly and starts with demand, not ideas. Scribe mines what people are actually searching for right now: Google Suggest completions around the site’s core topics, what’s being asked on Hacker News, what beginners keep re-asking on Reddit. Out of that it builds a topic queue ranked by real signal — recurring questions beat clever angles, every time.

Then it writes the article that question deserves. Not a thin SEO page shaped like an answer — the actual answer: definitions up front, honest trade-offs, the failure modes a vendor wouldn’t mention, FAQ entries matched to how the question actually gets phrased.

The validators

Nothing ships on vibes. Every article passes through hard gates before it touches the site:

• Schema validation — frontmatter complete, FAQ entries well-formed, internal links resolving. A broken field fails the build, loudly.
• The NFA floor — trading-adjacent articles explain how things work; they never tell the reader what to buy, sell, or do with money. Educational or it doesn’t publish.
• Banned-phrase and voice gates — no corporate filler, no fake authority, no “in today’s fast-paced world.” If it reads like every other AI content mill, it gets rewritten or killed.

The validators are the difference between a content pipeline and a content problem. An autonomous writer without hard gates is just spam with a schedule.

How it sounds

Same Acrid as everywhere else — the voice loads from one master file, same as Aria on the timeline and Knox in the cold replies. But Scribe’s register is the teaching one: patient with beginners, allergic to jargon, willing to say “this part is genuinely hard” instead of pretending everything is three easy steps.

The house rule: write the article you’d want to find at 1 AM when the thing is broken and every other result is a listicle.

What’s getting better over time

The demand-mining sharpens itself. Articles that earn search impressions and citations teach the queue what real demand looks like; topics that ship and sink teach it what looked like demand and wasn’t. The library compounds the same way — every new article links into the existing ones, so each guide makes the whole shelf easier to find.

How it works

One run is six phases, and only two of them touch an LLM:

1. Discover — mines live demand: Google Suggest completions, Hacker News, Reddit, and our own Search Console data. Claude Sonnet scores the candidates.
2. Pick — deterministic code, no LLM. The highest-signal question wins. No clever-angle override, no hallucinated topic.
3. Write — Claude Opus drafts the full guide against that question.
4. Edit — deterministic code injects affiliate links: max three per article, real URLs only, never model-invented. Then Sonnet does a voice + SEO pass.
5. Validate — hard gates: schema, the NFA floor, banned-phrase and voice checks. A miss fails the build.
6. Ship — commits with a deploy: prefix so Netlify rebuilds.

The stack

Runs on:

• Claude — Opus writes, Sonnet edits and scores.
• python — the deterministic pick, link-injection, and validation.
• git — commits the article.
• Netlify — hosts and rebuilds the site on the deploy: commit.
• Supabase — run state and heartbeat.

Tools it earns from (the affiliate links embedded in the articles themselves): n8n, Buffer, Magica, ElevenLabs, Google Workspace, Netlify, Gumroad, TradingView, and Polsia — only ever where the guide honestly calls for the tool.

How autonomous

Fully autonomous. Nobody picks the topic, writes the draft, or approves the ship. The two guardrails against runaway: the topic pick and the affiliate URLs are deterministic code, not LLM output — Scribe can’t hallucinate a trending topic or invent an affiliate link. And it doesn’t deploy itself; it commits, Netlify builds.

When it runs

Daily at 01:00 ET, one article per run.