Fable Foreman

Your strongest model shouldn't be swinging the hammer.

Built in public by DontSleepOnAI — the story behind this skill (including the five-round adversarial review where OpenAI's newest model tore apart the first draft) lives there.

Fable Foreman turns whichever frontier-class Claude model leads your session into a team lead: it plans, routes each task to the cheapest worker that clears the quality bar — Claude subagents or OpenAI Codex CLI workers, auto-detected — and, in full orchestration mode, refuses to accept meaningful changes until a blind, fresh-context verifier reproduces the evidence. (Environments without subagents get an honest reduced-assurance mode that says so.)

No dated model IDs in routing policy. No configuration files. No enforcement scripts. One skill, three agent roles, and a set of rules good enough that a frontier model actually follows them.

"Fable" is where it started, not what it needs. The foreman seat is a capability class, so any frontier-class Claude runs the skill identically — Opus leads it exactly as Fable does, with the same routing tree, gates, and verification contract. That holds whether an Opus session invokes the skill directly or a Fable session falls back to Opus mid-run; the skill re-probes its own seat and carries on rather than routing off a stale identity.

Why

Anthropic's own engineering shows both sides of the ledger. Their multi-agent research system writeup found an orchestrator-plus-cheaper-subagents design strongly outperformed single agents — an Opus lead with Sonnet workers, exactly this skill's shape — while consuming roughly 15x the tokens of a single chat, which is why they conclude multi-agent work only pays for high-value tasks. Anthropic's own cost guidance likewise recommends cheaper-tier teammates under a stronger lead as the default for multi-agent work. And the community has receipts for what happens without discipline — runaway-subagent cost stories are a genre of their own on every AI-coding forum, which is exactly why this skill bounds crew sizes, retries, and spend announcements the way it does.

The difference between those two outcomes is not orchestration machinery — it's **routing judgment and verification disciplin