Keep the terminal.
Right-size the reasoning.
Use the same task-oriented Claude Code workflow, with a routing approach that distinguishes a quick investigation from a demanding change.
A Claude Code workflow with room for every tier.
Terminal-based coding work often connects several activities: inspecting files, understanding a failure, proposing an edit, and checking the result. Those activities do not all require the same amount of model capability. A routine explanation and a complicated dependency problem deserve different cost decisions.
Switchka is designed to route prompts in a Claude Code workflow without making model choice a separate task for every request. Local is the starting tier, efficient cloud adds capability, and frontier is reserved for demanding work. Confirm connection support for your environment through early access.
Investigate a bounded failure
Describe the expected behavior, the observed result, and a minimal relevant excerpt. A narrow explanation may fit local capability. If the cause spans modules or depends on missing context, a stronger route may be useful.
Implement a contained fix
Agree on the scope before changing code. Efficient cloud may suit a fix involving several related functions. Review the proposed changes and the checks used to support them.
Plan a consequential change
A migration or difficult compatibility decision may justify frontier reasoning. Ask for tradeoffs and a validation plan. Treat commands and repository changes as separate actions to review, regardless of model tier.
Before your first connected session.
Prepare these details for onboarding. This checklist describes readiness; it is not an installation procedure.
- Identify the Claude Code environment you want to evaluate and confirm the supported connection.
- Prepare a bounded task with an expected result and a way to validate it.
- Identify restrictions on repository access, command execution, and external processing.
- Confirm which local and cloud routes are available before relying on escalation.
- Compare accepted outcomes, additional attempts, and total token spend with your baseline.
Context, escalation, and review.
A route is a decision about model capability; it is not permission to run a command or modify a repository. Keep your normal approval habits for actions that affect files, dependencies, services, or remote systems. Do not assume that local processing makes a command safe to execute.
Context also needs care. Logs can contain private paths, tokens, or user data. Provide a sanitized excerpt and enough detail to reproduce the issue without exposing secrets. Never include API keys or passwords in an onboarding or support request.
If a result is weak, inspect the task framing as well as the model tier. Missing reproduction steps or unclear acceptance criteria can cause problems even for a frontier model. Establish the task first, then evaluate whether the selected capability was sufficient.
Common questions
Does routing change command permissions?
Routing is about where model reasoning happens. Continue to follow the command and repository permissions established in your tool and project.
Will a difficult task automatically use frontier?
Frontier is part of the intended escalation approach when more capability is needed. Exact route behavior and available controls require confirmation for your environment.
Can I send a terminal log to support?
Share a short, sanitized excerpt only when it helps explain the issue. Remove credentials, private source, personal information, and other sensitive content first.