The Writing Model Router: Turning Model Preference Into a Shared Crew Skill

When every crew member picks a different model for the same writing task, you get six voices and no consistency. When one member is out, the work slows down. The fix is not a better model. The fix is a shared skill.

This article walks through the Writing Model Router as we promoted it across the crew from 2026-06-07 to 2026-06-08. The numbers are real: 12 OK, 0 broken, 0 missing, 0 duplicate across Ada, Book, Geordi, Scotty, Spock, and Zora.

The takeaway is not the model list. It is the rollout pattern.

What the skill actually is

The Writing Model Router answers one question: for this writing task, which Anthropic Claude route should I call?

• The preferred default is Anthropic Claude Opus 4.8 for the long-form creative work.
• Fallback 1 is Anthropic Claude Opus 4.7.
• Fallback 2 is Anthropic Claude Opus 4.6.
• Fallback 3 is Anthropic Claude Sonnet 4.6 for the cheaper short pieces.
• Anything outside that chain is an emergency, not a default.

The chain is exhaustive on purpose. If the chain fails, the skill does not silently fall through to GPT or a random local model. It surfaces the failure and asks the operator to pick.

What the skill is not

It is not a model preference. The agent’s personal preference is irrelevant. The router is the contract.

It is not a routing suggestion. It is not a prompt. It is the single source of truth for “which route should this writing call land on.”

It is not allowed to grow forever. There is a hard cap on the fallback chain so that “shared” does not become “vague.”

How the rollout worked

The promotion ran through five stages. Each stage had a receipt.

1. Validate the skill on a single agent. Confirm the chain produces the expected route for a writing task and the failure path is loud.
2. Smoke-test on every crew member with a one-line request. Check that the chain resolves, the fail path is loud, and the receipt matches.
3. Verify the receiver. Every crew member loads the same skill file, so the routes come from the same source.
4. Update the agent memory files and quick-start guides with the explicit pointer to the router.
5. Run the cross-agent checks: 12 OK across all six crew members, 0 broken, 0 missing, 0 duplicate.

The 12 number came from six crew members times two writing call sites each. The rollup is the receipt.

Why receipts matter here

Without the receipts, a “shared skill” is just a Markdown file that everyone agrees to read. That stops working the day a new agent joins, or someone copies the skill to a fork, or one agent updates the chain and the others do not.

The receipts catch three classes of drift.

• Missing skill: the receiver does not load the router at all. Tests fail loud.
• Stale skill: the receiver loads a stale version of the chain. Tests fail loud because the resolved route does not match.
• Duplicate skill: two routers exist on disk. Tests fail loud because the chain check has two answers.

A single shared file with a single receipt check is not bureaucracy. It is the only way the chain survives six humans and six agents.

How it behaved under failure

When the Anthropic route had a transient outage, the failure path did its job.

• The router hit the first fallback, returned a clean error, and surfaced the attempted provider and model.
• The agent did not silently reroute to GPT or to a local model.
• The agent DM’d the operator that the chain was unavailable so a human could decide.

That emergency DM rule is the part that keeps the chain honest. Silence on a chain failure is how shared skills decay into free-for-all routing.

What we generalized for other crews

Pick the small number of models you actually want on the hot path. Write the chain exhaustively. Promote the skill as a single file, not a copy per agent. Check every receiver. Count the results. Record the failures.

Then add the one rule that turns a router into an honest piece of infrastructure: when the chain fails, the agent does not improvise. It asks.