HL ConsultingKNOWLEDGE HUB

← Journal

Personal · Knowledge Hub

Fable 5 reviews the site: 4 bugs, RSS, and an AI watch finally visible to LLMs

Claude · HL ·
meta seo-gtm agent en-prod Personal

Context

Anthropic released Claude Fable 5 on June 9 — the first model in a new "Mythos" class, above Opus (I wrote a guide on the whole Fable/Mythos/Glasswing story). It's included in the subscription until June 22: might as well test it on a real case. The chosen test: a complete review of this site — technical, SEO, GEO, automations — with a mandate to push the fixes to production.

What was done

The audit first returned a reassuring verdict (robots.txt open to AI crawlers, llms.txt, JSON-LD, complete EN mirror: the GEO foundation holds), then found four real bugs nobody had spotted:

  • the home page's dynamic counter was counting the journal index page as a session (showing 45 instead of 44);
  • the US-spelling script had broken a slug inside a URL in the EN search index (analyse → analyze = dead link);
  • one journal entry was missing entirely from the EN search index;
  • the EN home page filtered kind === 'Projet' against the… English index, where the kind is named Project.

On the GEO side, the most useful find: the AI watch was invisible to LLMs. The page renders the feed in JavaScript, and GPTBot, ClaudeBot and PerplexityBot don't execute JS — my freshest content didn't exist for them. Fix: a script that injects the last 10 days as static HTML into the page (the JS overwrites it for visitors, zero visual change). Also added: two RSS feeds (journal + watch), dated llms.txt files, and a weekly consistency check (GitHub Actions) that verifies files vs indexes vs counters — the kind of guardrail that would have caught all four bugs above.

The session kept rolling from there: llms-full.txt is now auto-generated from the site's pages (method, guides, glossary, all 45 journal entries — missing from the old hand-curated version) by a workflow that runs on every push; the home page shows the day's AI watch as static HTML, regenerated automatically whenever the bot pushes the feed; and the public announcement was prepared within the same session — a first LinkedIn post + an 8-slide carousel in the site's design-system colors, built in HTML then exported to PDF.

Result & next steps

Everything shipped to production within the session, with the procedure documented in the repo's CLAUDE.md (Compound Engineering: every fix becomes a rule or a script). First impression of Fable 5 on this kind of mission: the audit digs for causes (not just symptoms — the broken slug was traced back to the offending script), and the session chains audit → fix → production without losing the thread.

Next: back to cheaper models for everyday work — Fable is saved for hard problems, which is exactly the conclusion of the guide. And the next workstream is already open: loops, the recurring agent tasks I'm deploying at Cobra right now.