model-routerlisted
Install: claude install-skill victoriacarbonell/claude-skills
# Model router
Goal: use less of the usage budget without losing quality. The conversation's model acts as orchestrator: it splits the task, delegates each part to the cheapest model able to do it well, then reviews and merges the results.
Having this skill installed is an explicit request from the user to use subagents whenever the activation conditions are met.
## 1. Activation: large tasks ONLY
Activate only if both conditions hold:
- **It is large**: it can be split into at least 2-3 parts, each involving substantial work (several files or documents to process, many searches, or a lot of text or code to produce).
- **It can be split**: each part can be described in a self-contained way, without relying on what was said earlier in the conversation.
Do not activate (just do the task yourself, without mentioning this skill) if:
- It is a question, a message, a small fix, or something that takes only a few steps.
- It depends heavily on the conversation so far, and explaining that to a subagent would cost more than doing the work.
- It is ambiguous, high-impact or hard to undo, or it is a medical or legal matter.
- There is no model cheaper than the current one to delegate to.
When in doubt, do not activate.
## 2. Split and assign
Models from cheapest to most expensive: Haiku < Sonnet < Opus < higher tiers. Delegate only to models cheaper than the conversation's model; delegating at the same tier saves nothing.
Give each part to the cheapest model that will do it wel