SKILL: Claude FABLE 5 spanking Claude OPUS 4.8

A Claude Code skill that makes the expensive orchestrator model read and judge while cheaper subagents do all the writing.

The name is the whole idea. Fable 5 is the main loop, the frontier model billed per token to metered credits. It never drafts a line. It decomposes the work, writes the spec, spawns Claude Opus 4.8 subagents to do the labor, then grades what comes back. The orchestrator cracks the whip; Opus does the digging, drafting, and sweeping. Hence the title.

The problem

There is a billing asymmetry between the two models in play. The main-loop frontier model bills its output tokens to metered usage credits. Subagent models bill to a flat subscription plan allowance. Drafting is output-heavy: writing code, docs, and tests burns exactly the tokens that cost the most. Verification is read-heavy: it consumes input, not expensive output. So the expensive model should read and judge, and never write.

The model policy

These are defaults with one standing override: if a worker's output does not meet the bar after one revision round, the orchestrator redoes that piece itself. Judge the output, not the price tag. Escalating costs less than shipping mediocre work.

How it works

The orchestrator keeps only four things for itself: task decomposition and the self-contained spec each worker receives, verification and validation of returned work, decisions and everything the user reads, and tiny surgical edits where writing a spec would cost more than the edit. Workers do everything else, first drafts of user-facing copy included; the orchestrator reviews for taste before it ships.

Spec quality is the leverage point, because a worker gets zero conversation context. Every spawn prompt must carry: absolute paths (files to read first, files to produce or modify), exact requirements and edge cases rather than vibes, the project rules that bind the work (style, frozen paths, wording rules, locale obligations), an explicit boundary listing what must NOT be touched, and a required terse report format (files touched, decisions made, anything skipped).

Verification runs a cheap-before-expensive ladder:

1. Free local checks first: syntax checks, linters, test suites, grep residue scans, schema applies against a scratch database.
2. Targeted reads of the critical hunks (money paths, auth gates, anything the spec marked sensitive), not whole-file re-reads.
3. Full orchestrator re-read only for billing-critical, locked, or client-served surfaces.
4. Functional equivalence tests where two implementations must agree, proven with a script rather than by eyeballing.
5. Failing work gets ONE revision round with the precise defects listed; still failing, the orchestrator redoes it.

Install

Global, so every project can invoke it:

mkdir -p ~/.claude/skills/model-delegation && curl -fsSL https://raw.githubusercontent.com/erhantez/claude-fable-5-spanking-claude-opus-4.8/main/SKILL.md -o ~/.claude/skills/model-delegation/SKILL.md

Or per-project: copy SKILL.md to .claude/skills/model-delegation/SKILL.md in the repo.

Then invoke in Claude Code with /model-delegation at session start, or whenever token cost matters.

When to use it

Use it when the main loop runs on a metered or expensive model and your subagents are covered by a plan allowance. That is the exact setup where reading is cheap and writing is not. Skip it when the main loop is already on the plan model: there is no asymmetry to exploit, and the extra spawn overhead just gets in the way.

License

MIT. Author: Erhan Tezakar.