How I Direct AI: A Real Case

Published 2026-07-09

Background: Claude Code's top model, included with existing paid subscriptions until July 12

In July 2026, Anthropic opened its new flagship model, Fable 5, to paying subscribers, but only until July 12. A flagship model burns tokens fast, and spending it on trivial tasks that require no real reasoning is a waste.

To conserve tokens, I gave Claude Code a very simple instruction:

The AI did not comply; it audited my design first

The AI's answer (excerpted for readability):

This is a meta architecture-design question (not a tax or technical domain question, so no project files needed; the basis is this harness's actual tool specifications plus your CLAUDE.md and memory). Overall verdict first, then the technical traps and the corrections.

Overall verdict

The direction is right: this is the orchestrator-worker pattern, and it naturally complements your existing multi-model blind-review rule: Fable × Opus is genuine cross-model review, which catches more blind spots than any model grading its own work. But a literal implementation would hit four technical traps that need fixing first.

1. Delegated subtasks default to running on the main model (meaning a delegation that forgets to specify the executing model still burns flagship quota, so the delegation achieves nothing); every delegation must name its model explicitly.
2. The "scale effort to complexity" parameter I had imagined does not exist on the single-delegation tool at all; it requires a separate batch-orchestration path.
3. A subtask is a brand-new conversation that sees none of the main thread's context, so every delegation needs a self-contained brief; for small tasks, the communication overhead costs more than the main model just doing the work.
4. Framing the task, grading the output, and arbitrating disputes all consume flagship quota too: delegation saves the "execution", never the "judgment".

The key correction: from "by importance" to "by leverage of judgment"

The AI then went after the core of my design:

The suggested routing table (corrected)

Your split is "by importance"; I suggest "by leverage of judgment": important-but-mechanical work still gets pushed down, and Fable's judgment goes to the two ends, the spec and the review, not to the typing in between.

The corrected layering:

With three supporting rules:

• Delegation threshold: the task is self-contained, the output compresses into a summary, and no mid-course judgment is needed. All three must hold before delegating. Handing work off feels efficient, but because a subtask sees none of the main thread's context, an unclear brief makes the rework cost higher, so the main model is better off doing it directly.
• Escalation triggers: on any irreversible action, conflicting sources, or low confidence, the delegated task stops and reports back rather than pushing on.
• Cross-model review: the second-tier model's output is reviewed by the flagship; conversely, the flagship's own output goes to a different model for independent blind review (a model grading its own work does not count, a rule I have kept for years).

Checking the same answer from a second angle

The same design can be verified from another direction. From the quota angle, concentrating judgment at the spec and review ends while delegating execution maximizes the yield of a scarce resource. From the quality angle, cross-model review lets two different models catch each other's blind spots, one more layer of protection than any model reviewing itself. Both angles point to the same conclusion, and only then was the corrected version adopted.

The final review and sign-off stays with me

Across the whole exchange, the AI did two things: it exposed the gap between my design and the system's reality, and it proposed the correction. But some things could not be left to the AI to execute. Where important work with an already-clear spec belongs was my decision: second-tier model drafts, flagship model reviews. Whether to adopt the whole mode, and how far its authorization runs, was also my decision.

This is the concrete version of the line on my About and Services pages: the code and the text can come from AI; the value lies in designing the process and verifying the output. Directing AI is not a prompting trick. You need your own standard of judgment first, so the AI's pushback has something to be tested against, and the final call has something to rest on.