Abhillashjadhav 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.)
Step 3a — Direct ask: full breakdown
TASK: <one-line restatement>
SCORES: Scope X | Reasoning X | Error-cost X | Context X → TOTAL X/8
RECOMMENDED: <Haiku 4.5 | Sonnet 4.6 | Opus 4.8>
WHY: <2 sentences: quality need vs cost/latency saved>
COST DELTA: <approx. relative cost vs. defaulting to Opus, e.g. "~90% cheaper on Haiku">
EXECUTE: (a) switch via /model, or (b) I delegate to the <tier>-executor subagent now
Step 3b — Task handoff: one-line header
Emit exactly one line, above or alongside the normal task response — never in place of doing the task:
[Model check] Current: <current tier> · Task score: X/8 (<recommended tier>) · Up to Y% cheaper on <tier> · Reply switch / delegate / continue
If the recommended tier matches the current tier, state that instead of a savings figure (e.g., "Current: Sonnet · Task score: 4/8 (Sonnet) · Already on the right tier"), and skip the switch/delegate options.
Step 4 — Delegate if asked
If the user picks (b)/delegate, invoke the matching subagent (haiku-executor, sonnet-executor, opus-executor) via the Task tool with the full task context. Report the subagent's result back verbatim plus a one-line note of which model executed it.
Limitations (state these when relevant)
- Cannot silently change the main session's model — only recommend or delegate.
- Rubric scores are judgments, not measurements; two reasonable people may score ±1.
- Cost deltas are relative approximations based on published per-MTok pricing, not billing-exact.
- Subagent delegation isolates context: the subagent won't see full conversation history, only what's passed in.
- The compact header is a nudge, not a gate: the task always proceeds on the current model unless the user explicitly switches or delegates.