skill
Research, compare, and select LLM models for AI-powered apps and workflows. Finds latest models per family, verifies availability on target platform, compares pricing/benchmarks/tool-calling, produces ranked recommendations. Use when: "which model", "compare models", "find a model", "model research", "best model for", "cheapest model", "tool calling models", "model selection", "upgrade model", "swap model", "fallback models", "model chain". Trigger: /model-research.
/model-research
Source: skills/model-research/SKILL.md
Research, compare, and select LLM models for AI-powered apps and workflows. Finds latest models per family, verifies availability on target platform, compares pricing/benchmarks/tool-calling, produces ranked recommendations. Use when: "which model", "compare models", "find a model", "model research", "best model for", "cheapest model", "tool calling models", "model selection", "upgrade model", "swap model", "fallback models", "model chain". Trigger: /model-research.
Support primitive in the Harness Kit operating loop.
This generated excerpt gives readers the beginning of the live primitive contract before they jump to GitHub.
Find the right model for a specific use case. Live data, not training data.
## The Iron Laws
```
1. NEVER RECOMMEND MODELS FROM TRAINING DATA ALONE.
Every model name, price, and benchmark MUST come from a live source.
2. EVERY MODEL HAS A RELEASE DATE. USE IT.
The current date matters. A model released 3+ months ago may already be
superseded. A model released last week may be unstable. Factor age into
every recommendation.
```
LLM model landscapes change weekly. Training data is stale within months.
A model recommended from memory may be discontinued, superseded, or repriced.
## Delegation Floor
Delegation floor applies: probe the roster first; dispatch two or more
providers for substantive model selection; direct solo only for mechanical
catalog fetches, emergency preservation, user-forbidden delegation, or
fewer-than-two-providers cases. See `harnesses/shared/AGENTS.md` (Roster).
Local lane guidance: Use specialized lanes for platform catalog verification,
benchmark/source review, social/reliability pulse, integration/codebase fit,
and cost/latency risk. Prefer different model/provider families for independent
judgment; the lead owns synthesis and final recommendation.
### The Freshness Rule
**Today's date is always available in system context. Use it.**
For every model under consideration, determine its release date and compute age:
| Age | Classification | Implication |
|-----|---------------|-------------|
| < 2 weeks | **Bleeding edge** | May have bugs, API instability, sparse benchmarks. Flag risk. |
| 2 weeks – 2 months | **Fresh** | Sweet spot: stabilized but still current-gen. Prefer these. |
| 2 – 4 months | **Aging** | Check if a successor exists in the same family. Often superseded. |
| 4+ months | **Stale** | Almost certainly superseded. Do NOT recommend without verifying it's still the latest in its family. |
...