Workflow: Model Routing

Principle

Use the cheapest model that is still strong enough for the current phase.

The stack documented here separates everyday execution from deep reasoning on purpose.

Agent	Model	Typical use
`build`	`openai/gpt-5.3-codex`	default implementation, local coding, fast iteration
`plan`	`github-copilot/claude-sonnet-4.6`	normal planning, analysis, architecture discussion
`build-deep`	`openai/gpt-5.4`	harder implementation, heavier reasoning, risky edits
`plan-deep`	`github-copilot/claude-opus-4.6`	requirement clarification, difficult planning, high-quality writing

Subagent	Model	Role
`explore`	`github-copilot/gemini-3-flash-preview`	fast read-only exploration
`general`	`github-copilot/gpt-5-mini`	cheap general helper tasks
`reviewer`	`github-copilot/claude-sonnet-4.6`	review quality and risk

Command	Model strategy	Why
`/plan`	`claude-opus-4.6`	interactive planning and stronger written output
`/spec`	`gpt-5.4`	turn acceptance criteria into clean test skeletons
`/implement`	`gpt-5.3-codex`	default coding workhorse
`/review`	reviewer subagent	review precision and consistency
`/verify`	`gpt-5.4`	better failure diagnosis

stay on build for most coding
switch to build-deep only when the implementation is genuinely hard
use plan for normal design conversations
use plan-deep or /plan when the problem is still fuzzy, risky, or user-facing text quality matters

This keeps the stack fast by default while preserving access to stronger reasoning when it actually pays off.

Always verify the exact model IDs available from your providers before copying the templates verbatim.