model-complexity-router

Model Complexity Router

Classify a concrete task, recommend the right Claude model tier, and optionally delegate execution to a subagent pinned to that tier.

Why this exists

The main session's model cannot be programmatically switched — no hook, skill, or API changes it mid-session. What IS possible: (a) a documented recommendation the user acts on via /model, or (b) delegation to a subagent whose frontmatter pins a model. This skill does both and says so honestly.

Step 1 — Detect the trigger mode

Every time the user hands over a substantial task — a concrete thing to build, fix, write, or execute — this fires silently in the background, in parallel with doing the task. It never blocks or delays the response.

Two modes:

• Direct ask ("which model should I use", "is this an Opus task", etc.) → full scored breakdown (Step 3a).
• Task handoff (any other substantial task, proactively) → silent score, one compact line (Step 3b).

If the user gave no concrete task at all (e.g., "which model is best generally?"), stop and ask for the specific task — never classify hypotheticals. Fire once per distinct task: once a line or block has been shown for a task, do not re-emit on every subsequent message about that same task — only when a new, distinct task is handed over.

Step 2 — Score against the rubric

Score the task on four dimensions, each 0–2:

Dimension	0	1	2
Scope	Single file/artifact, cosmetic	One feature, few files	Multi-file, architecture, cross-system
Reasoning depth	Mechanical/pattern-match	Some judgment, known patterns	Novel tradeoffs, ambiguity, design decisions
Error cost	Trivially reversible	Rework hours	Wrong answer compounds (prod, money, public)
Context load	Fits in one prompt	Needs a few files	Needs large context synthesis

Total 0–8. Map: 0–2 → Haiku, 3–5 → Sonnet, 6–8 → Opus.

Floor rule: if Error cost = 2, the recommendation can never be Haiku — floor at Sonnet. (Error cost is already in the total; do not bump the tier a second time.)