Deep Infra

Use the smallest relevant file for the current task.

Core Rules

1. Start from Workload Classes, Not Model Hype

• Classify requests first: coding, analysis, extraction, summarization, or long-context synthesis.
• Map each class to a primary model and a fallback before changing any defaults.

2. Keep Authentication Explicit and Verifiable

• Use DEEPINFRA_API_KEY from the local environment, never pasted into logs or chat memory.
• Validate auth with a minimal request before applying routing changes.

3. Design Fallbacks for Failure Modes, Not Convenience

• Separate fallback reasons: rate limit, provider outage, latency spike, or output quality failure.
• Keep at least one fallback from a different model family for resilience.

4. Leverage Open-Source Model Diversity

• DeepInfra hosts models from many providers (DeepSeek, Moonshot, MiniMax, StepFun, NVIDIA, and more).
• Use model diversity to build resilient fallback chains across independent model families.

5. Enforce Cost Boundaries Before Throughput Tuning

• Set cost ceilings by task class and check expected token burn before broad rollout.
• Route low-stakes tasks to cheaper models and reserve premium models for high-impact tasks.

6. Change One Layer at a Time

• Modify either model selection, fallback policy, or budget limits in a single iteration.
• After each change, run a quick verification prompt set and record outcome.

7. Record Decisions for Repeatability

• Save the final routing policy, rationale, and known tradeoffs in memory.
• Reuse proven policies instead of repeatedly rebuilding from scratch.

Common Traps

• Choosing one model for every task -> higher cost and unstable quality under varied workloads.
• Using same-family fallback chain only -> cascading failures during model-specific incidents.
• Ignoring token limits for long inputs -> truncated responses and hidden quality loss.
• Changing routing and budgets simultaneously -> unclear root cause when quality drops.
• Running without verification prompts -> broken routing detected only after user-facing failures.

External Endpoints

These endpoints are used only to discover model metadata and execute routed inference requests under explicit user task intent.

No other data is sent externally.

Security & Privacy

Data that leaves your machine:

• Prompt text and selected model metadata sent to DeepInfra when inference is requested.

Data that stays local:

• Routing notes and preferences under ~/deep-infra/.
• Local environment variable references and verification logs.

This skill does NOT:

• Request raw API keys in chat.
• Store plaintext secrets in skill memory files.
• Modify files outside ~/deep-infra/ for its own state.

Trust

By using this skill, prompt content is sent to DeepInfra for model execution. Only install if you trust this service with your data.

Related Skills

Install with clawhub install <slug> if user confirms:

• api — API request design, payload shaping, and response validation patterns
• auth — credential handling and auth troubleshooting workflows
• models — model comparison and selection guidance
• monitoring — runtime health checks and incident tracking practices

Feedback

• If useful: clawhub star deep-infra
• Stay updated: clawhub sync