name	shipyard-cp-local-routing
description	Use when operating shipyard-cp and deciding how to mix local Qwen 4B or 27B runtimes with external providers. Apply it when the goal is to save tokens, move prep work to local models, reserve premium providers for high-risk stages, or define model roles across plan, dev, acceptance, integrate, and publish.

shipyard-cp Local Routing

Use this skill to decide where local models fit into a shipyard-cp workflow.

Use local Qwen3.5-4B for routing, summarization, classification, compression, and checklist extraction.
Use local Qwen3.5-27B for rewrites, draft generation, and first-pass review where local quality is good enough.
Use external frontier providers for high-risk code changes, final acceptance judgment, and integrate or publish gates.
Keep final gate evidence in shipyard-cp task, run, and audit logs even when local models are used in earlier stages.

Split each task into cheap prep work versus quality-critical output.
Push cheap prep work to local models first.
Send only the reduced, relevant context to premium providers.
Prefer local 4B before local 27B when the work is mostly structured or classification-like.
Prefer local 27B before an external provider when a local draft can reduce expensive tokens later.

For runtime control, use the shared launcher skill and commands:

name	shipyard-cp-local-routing
description	Use when operating shipyard-cp and deciding how to mix local Qwen 4B or 27B runtimes with external providers. Apply it when the goal is to save tokens, move prep work to local models, reserve premium providers for high-risk stages, or define model roles across plan, dev, acceptance, integrate, and publish.

Use this skill to decide where local models fit into a shipyard-cp workflow.

Use local Qwen3.5-4B for routing, summarization, classification, compression, and checklist extraction.
Use local Qwen3.5-27B for rewrites, draft generation, and first-pass review where local quality is good enough.
Use external frontier providers for high-risk code changes, final acceptance judgment, and integrate or publish gates.
Keep final gate evidence in shipyard-cp task, run, and audit logs even when local models are used in earlier stages.

Split each task into cheap prep work versus quality-critical output.
Push cheap prep work to local models first.
Send only the reduced, relevant context to premium providers.
Prefer local 4B before local 27B when the work is mostly structured or classification-like.
Prefer local 27B before an external provider when a local draft can reduce expensive tokens later.

For runtime control, use the shared launcher skill and commands:

shipyard-cp-local-routing