ModelSense Skill

Purpose

ModelSense helps users pick the optimal model and effort level for their task.
It does NOT route automatically on every request (use a provider plugin for that).
It's an on-demand advisor: ask it a question, get a clear recommendation with reasoning.

When to trigger

• - User asks: "which model for X?", "should I use Opus or Sonnet?", "what effort level?"
• User wants to understand what a benchmark means
• User wants ModelSense to auto-switch the session model

Inputs to collect (infer from context, ask only if truly unclear)

1. 1. Task description — what is the user trying to do?
2. Effort preference (optional): quick / balanced / deep / INLINECODE3

- If not specified, infer from task urgency/complexity

1. 3. Auto-switch? — does the user want ModelSense to apply the recommendation automatically?

Recommendation Process

Step 1 — Task Analysis

Classify the task across these dimensions:

• - Domain: code, math, reasoning, writing, dialogue, document analysis, multimodal, research
• Complexity: simple / moderate / complex / research-grade
• Output type: text, code, JSON, long-form, structured data
• Context length needed: short (<8K), medium (8–32K), long (32K+), very long (100K+)
• Special requirements: function calling, thinking/CoT, multimodal, speed-sensitive

Step 2 — Benchmark Matching

Cross-reference task domain with relevant benchmarks from data/benchmarks.yaml.

Benchmark	Best for
HumanEval / SWE-bench	Code generation, debugging, engineering
GPQA

Graduate-level science & research |
| MATH / AIME | Mathematical reasoning |
| MMLU | General knowledge, multidomain QA |
| Needle-in-Haystack | Long-context retrieval |
| MT-Bench / Arena Elo | Dialogue, writing quality |
| BBH (Big-Bench Hard) | Complex reasoning, multi-step logic |

Step 3 — Effort × Model Matrix

Effort	Target quality	Typical model tier
INLINECODE5	Good enough, fast	Haiku / Flash / GLM
INLINECODE6

High quality, reasonable cost | Sonnet / GPT-4o | | deep | Best available, thinking on | Opus / o3 | | research | No cost limit, maximum quality | Opus + thinking=high |

Step 4 — Provider Filter

Check the user's available providers:

• - Run: openclaw models list via exec tool (or read from context)
• Only recommend models the user can actually use
• Flag when a top pick requires a provider they haven't configured

Step 5 — Output the Recommendation

Format:
CODEBLOCK0

Auto-Switch Behaviors

Option A: Advisory only (default)

Just output the recommendation. Tell user: "Run /model <name> to switch."

Option B: Switch current session

If user confirms or says "yes switch" / "apply it":

session_status(model="<provider/model>")

Notify user: "✅ Switched to X for this session. Run /model default to reset."

Option C: Delegate task to best model

If user says "just do it with the best model": CODEBLOCK2

Data Files

• - data/benchmarks.yaml — benchmark definitions, score leaders, task mappings
• INLINECODE13 — model catalog (updated via GitHub Actions weekly)

Examples

User: "I need to write a Solidity audit report"
→ Domain: code + security + long-form
→ Benchmarks: SWE-bench, HumanEval
→ Recommendation: claude-opus-4-6 with thinking=high, effort=INLINECODE16

User: "Quick summary of this Slack thread"
→ Domain: dialogue, short
→ Recommendation: claude-haiku-4-5 or gemini-flash, effort=INLINECODE19

User: "Prove this mathematical conjecture"
→ Domain: math, research-grade
→ Benchmarks: MATH, AIME, GPQA
→ Recommendation: o3 or claude-opus-4-6 with thinking=high, effort=INLINECODE23