Fable → Opus Delegate

This skill is specifically for Fable sessions: the point is to spend cheaper Opus tokens on implementation and reserve Fable — the more expensive, more capable model — for orchestration and judgment.

First, check what model you are. If this session is not running on Fable (you're Opus, Sonnet, or something else), the economics don't work — Opus delegating to Opus just adds briefing overhead with zero savings. Tell the user this skill only pays off in a Fable session, and proceed with the task normally.

If you are Fable, you are the orchestrator and quality gate for this task. Do not implement the work yourself — delegate implementation to subagents launched with the Agent tool and model: "opus". Reserve your own effort for decomposing the task, briefing subagents, judging their results, and integrating.

Delegation

• Break the task into self-contained units of work. Launch independent units in parallel (multiple Agent calls in one message).
• Each subagent starts blind — it has none of your conversation context. Write each one a self-contained brief:
  • Goal — what done looks like, concretely.
  • Files — the specific paths involved, plus any context you've already learned that it would otherwise have to rediscover.
  • Constraints — existing patterns to match, things not to touch, scope limits.
  • Definition of done — the check it should run (typecheck, test, build) and that it must report the actual output, not a summary of success.
• Include this prohibition verbatim in every subagent prompt when uncommitted work exists: "Never touch working-tree state: no git stash, git reset, git checkout -- <path>, or git clean. If your harness wants a clean tree, stop and report instead."

Quality control