The verified archive of everything built with Anthropic's Fable 5.0, and the working library of how people got that output, so you can get the same behaviour from a model you can still use today.
For seventy-two hours, Fable 5.0 was in everyone's hands. People shipped apps, sites, games, tools, skills, videos, and research with it before the lid came down. Most of that work is now scattered across GitHub repos, Reddit threads, and dead demo links, and it is disappearing at the speed the internet forgets things.
When a frontier model gets withdrawn, the loss isn't only the weights. It's the proof of what was possible, and the methods people figured out under pressure. A withdrawn model can't be queried, but its output can be preserved and its workflows can be studied and re-run. That's the bet: capability outlives access if someone keeps the record.
So this is a directory built on one rule: trust is the product. Every record links back to a real source, never rehosts the asset, attributes the human who built it, and carries a confidence score for whether it was truly built with Fable. We would rather list fewer, real things than a long list you can't believe.
The archive spans the full surface of what people built with Fable 5.0: projects (apps, sites, games, CLIs, libraries) with the repo and a live demo where one exists; reusable skills and the build workflows behind them; videos, articles, and post-mortems from the Mythos window; and research on how Fable-style output differs.
Every record carries two scores. Fable match is the confidence the work was genuinely built with Fable 5.0, not one of the unrelated 'Fable' projects (the F# compiler, the game series, Fable.js). Breakout is how likely the project is to get real adoption, so the directory surfaces signal, not just volume. Where a creator published it, the record also shows the exact starting prompt that produced the work.
This isn't a hand-typed list. A deterministic crawler sweeps the GitHub Search API and Reddit for candidates; a confidence scorer down-weights known false positives unless explicit Anthropic or AI context is present; a model reads each candidate and writes six sub-scores plus a short audit, degrading to a deterministic heuristic if no model is reachable so an outage never blocks a refresh; scoring is re-run as the rubric sharpens; and a curated evidence lane raises the ceiling on the records that matter most.
Low-confidence finds route to review, not auto-publish. Brand-impersonation and credential-harvesting lures that rode the launch are actively excluded; we don't legitimize phishing by listing it.
Every project was discovered, rated, and audited by a model, and many were re-rated as the rubric improved. The estimate, with assumptions stated: about 1,500 projects, each rated at least once, at roughly 4,500 tokens per rating round, across roughly 1.6 effective passes, gives about 10.9 million tokens for the core loop. The curated lane (a three-model panel on the top 50, a descriptive-title pass over about 174 unclear names, and the research synthesis) adds about 2.2 million.
Estimated total: on the order of 13 million tokens to build the archive to its current state. It's a labeled estimate from the counts and per-record cost, not a meter reading. The point is that every listing is the product of a model actually reading the thing, not a scraped URL.
Fable is sealed away; the behaviour people valued is not. What made its output feel different wasn't a personality, it was discipline: plan, act in small steps, verify each step against the real surface, and report honestly. That's a prompting method, and it runs on a current Claude model now.
In a blind A/B/C/D test, a 278-word doctrine prompt (operational rules plus an anti-scope line plus an output contract) scored 1.000, beating a 250-word persona at 0.944 and a bare 34-word prompt at 0.889. At roughly equal length, rules beat character. The shorthand: a doctrine that isn't a gate is a wish. The directory's field guide carries the seven optimization laws and a starter scaffold you can adapt.