Use Claude Fable as orchestrator, architect, synthesizer, and final judge, and send the token-heavy research, coding, testing, and log reduction to cheaper subagents. “Fable should spend tokens on judgment, not on reading every file, reducing every log, or manually running every browser check.”
Fable keeps five jobs: decomposing ambiguous work into parallel slices, architecture and product tradeoffs, reading conflicting subagent reports, integrating partial implementations, and final review. Delegation is written as handoff packets, “as if the subagent has no useful chat context”. Each packet names the repo path, objective, in-scope and out-of-scope surfaces, the evidence format to return, verification commands, and stop conditions.
Vetting is not delegated. “Treat subagent reports as leads, not facts. Before using a high-impact finding, opening a PR, or telling the user the work is done, Fable should reopen the important cited files, confirm the relevant line refs or failures, and review the final diff.” The skill includes a performance claim (“up to 3-5x more cost-efficient and 2-4x faster” when work parallelizes) and labels it “workload-dependent estimates, not guarantees.”
Off-stage. One of the two capabilities that anchor pattern-model-tiering.
See Also
- builder-efficient-frontier — the model-agnostic generalization.
- builder-stay-within-limits — the budget pause that pairs with it.
- pattern-model-tiering — the pattern minted from this pair.