Parable: a Fable-style scaffold for Opus, built from head-to-head experiments

I spent a couple days running Fable and Opus head to head on the same real tasks to figure out where the quality gap actually comes from. While Fable will always be a fundamentally more intelligent model, the procedures in this scaffold can help close the gap, so I packaged them into a Claude Code skill set: https://github.com/SamBWarren/parable

It rests on five core ideas.

Disagreement is signal. For hard tasks, a director spawns an ensemble of firewalled reviewers who get identical prompts, and an adjudicator agent mines the contradictions between their reports. If two reviewers disagree, the adjudicator has to actually test the unknown. The best individual reviewer found 4 of 6 seeded bugs in my test, together the ensemble found all 6 (the same as Fable on High).

Rules decay, so re-inject them. A hook reminds agents of their most important instructions during long runs (at most once per 5 minutes of tool activity), so the skills they were given at spawn don't get buried under thousands of lines of context.

Naming a check is not running it. Agents have to prove rule compliance to their parent via pasted terminal output. I kept catching agents reciting a rule as if reciting were the same as following it. Now it's grep output or it didn't happen.

Conclusions in the channel, cognition in a thoughts file. Reasoning and dead ends go in a scratch file and the chat only gets verified conclusions. This is the part that makes the output actually read like Fable.

The ratchet. Every miss becomes a written rule the same hour, in the file where it would have been caught.

Usual caveat, this is n=1 on one codebase with me as the judge, which is why the testing harness ships with it so you can re-run everything yourself. Would appreciate anyone trying it and reporting back.