Adaptive Model — AI skill for Claude Code

Adaptive Model

A Claude Code skill that intelligently routes each step of your work to the optimal Claude model (Haiku, Sonnet, or Opus) based on what the task actually requires.

No more manually picking models. No rigid "Phase 1 → Phase 2 → Phase 3" workflows. Just continuous, contextual decisions.

What it does

Once activated at the start of a session, Adaptive Model stays on for the entire conversation. At every action — user message, internal step, sub-task — it evaluates three axes and picks the right model:

1. Cognitive complexity — how hard is this step?
2. Decision risk — how structural or reversible is it?
3. Current state — are we clarifying, executing, or stuck?

Examples:

Situation	Model	Why
User vaguely describes what they want	Haiku	Low complexity, just clarify
Choose between 3 possible architectures	Opus	Structural, high complexity
Plan validated, code the auth module	Sonnet	Clear execution
Sonnet failed twice on the same bug	Opus	Stuck, needs deep diagnosis
"Just add a console.log"	Sonnet	Trivial
Full security audit of the code	Opus	Expertise + global view

Key features

• Routes by capability, not by version name — it targets a tier (Haiku/Sonnet/Opus), and the tier aliases always resolve to the latest model, so new Claude models are adopted automatically with zero edits to the skill
• Transparent model switching — a discreet indicator shows when the model changes
• Automatic escalation — when Sonnet gets stuck, Opus is spawned with full context
• User override — say "use Opus" or "go fast" and the skill respects it
• No waste — never uses an overpowered model for a trivial task
• Every decision is independent — no predetermined sequence

Future-proof by design

The skill never routes by a hardcoded model version. It describes what a task *needs* (reasoning depth, speed, cost, stakes) and picks the tier that p