hedging8563/tokenlab-skills

TokenLab Cost Routing

Use this skill when a user asks how to reduce TokenLab cost, compare model prices, pick fallbacks, or route requests by quality, latency, and budget.

What this skill should deliver

• A compact routing recommendation with exact public TokenLab model IDs.
• A cost-aware fallback chain for the user's workload.
• A catalog/pricing lookup path that can be rerun.
• A note on which endpoint family each model should use.
• Guardrails for when not to switch models because doing so would change output, safety, or request semantics.

Preferred approach

1. Identify the workload and constraints:
   • chat, coding, agent loop, image, video, audio, embedding, rerank, translation, or multimodal
   • quality floor
   • latency target
   • budget or cost ceiling
   • native endpoint requirement
2. Read live public catalog signals before recommending:
   • GET https://api.tokenlab.sh/v1/models
   • GET https://api.tokenlab.sh/v1/models?recommended_for=<scene>
   • GET https://api.tokenlab.sh/v1/models/:model
   • GET https://api.tokenlab.sh/v1/models/:model/pricing
3. Build a chain with roles:
   • primary quality model
   • balanced default
   • fast fallback
   • budget fallback
4. If the user asks for exact cost, compute from live pricing and their estimated token/media volume. State units and assumptions.
5. For non-chat requests, inspect model details before changing parameters or endpoint family.

Output format

• One sentence stating workload and assumptions.
• A table with Route role, Model ID, Endpoint, Why, and When to fall back.
• One catalog command and one pricing command.
• A short implementation note for retries, rate limits, and user approval when quality would drop.

Avoid

• Do not invent prices, discounts, or model counts.
• Do not choose a cheaper model if that would silently remove required native behavior, tools, media support, safety constraints, or structured output guarantees.
• Do not expose TokenLab internal channel, physical provider, or routing details.
• Do not turn a user-provided model into a different model without saying why.
• Do not hardcode a fallback list without saying when it was checked or how to refresh it.

Edge Cases

• If catalog or pricing endpoints are unavailable, say that routing cannot be price-verified and provide only an example pattern.
• If the user asks for "cheapest", include capability and reliability tradeoffs.
• If billing risk is high, require explicit user approval before adding automatic fallback to paid media/video generation.