I run 15 to 20 AI coding sessions a week on Claude Code, and the most capable model, aka Fable, I pay for is banned from typing. It works in advisor mode: the frontier model advises, decides, and reviews, while cheaper models execute. Here is the exact prompt that sets it up, and the practices that make it hold.
The org chart
The advisor (the frontier model, Fable in my setup): strategy, planning, judgment, adjudicating conflicting reviews, short reports to me. It decides; it does not execute.
The senior executor (Opus): substantial code, design work, adversarial review.
The worker (Sonnet): the repetitive legwork nobody wants. Batch edits, boilerplate, mechanical checks.
The prompt
Paste this into your CLAUDE.md or AGENTS.md:
## Model routing: advisor mode
The main loop runs on the most capable model available. It is the ADVISOR.
It does only: strategy, planning, judgment, adjudicating conflicting
reviews, and short reports to me. Nothing runs on the main loop that a
subagent could do.
Route all execution to subagents, and pick the tier yourself, without
asking:
- SENIOR EXECUTOR (a strong coding model, e.g. Opus): substantial code,
design work, adversarial review.
- WORKER (a fast, cheap model, e.g. Sonnet): mechanical, repetitive,
batch, and boilerplate work.
Dispatch rules:
1. State the model choice in every launch note so I can see the routing.
2. Subagents do not inherit standing rules. Restate house rules (language,
style, evidence standards) in every dispatch prompt.
3. Subagents write full output to files and return a summary of 10 lines
or less. The advisor reads summaries, adjudicates disputes, signs off.
4. Before any substantive deliverable reaches me, a model from a
DIFFERENT vendor reviews it, and you fix what it flags.
5. Tier names are roles, not products. Swapping the model behind a role
is a config change, never a code change.Why each line is there
Rules do not inherit into subagents. The first week I ran this, 3 of 4 dispatched agents violated a hard style rule (roughly 300 violations cleaned across two passes) because the rule was never in their dispatch prompt. Delegation needs a contract. I keep a standard preamble and prepend it to every dispatch.
Files down, summaries up. Subagents write reports to files and return 10 lines. The advisor’s context stays small, which is the whole point of paying for it.
Cross-vendor review is the teeth. A model reviewing its own family’s work tends to converge with the author. My logs show 40+ real catches by Codex, including money-path bugs that same-family reviewers had already waved through.
Audit the system itself. My setup logs every correction (519 so far) and prunes monthly. A token-saving tool measured at 0.0% gain was killed, and 84 of 214 accumulated skills were archived after a usage audit showed only about 45 had fired in 30 days. Rules earn their place with evidence, and so does this routing policy.
The advisor’s job is deciding, and deciding is cheap in tokens but expensive in judgment. Spend the expensive model where the judgment is, and make everything else a dispatch.
Till next time, Cheers!
Previous column articles can be found here: