turbo_skill

Overview

This skill provides direct code generation via a hosted LLM (Cerebras) to rapidly produce multi-file implementations from a precise contract prompt. It integrates into a speed-run pipeline: write an exact contract, generate files, run tests, and apply small surgical fixes until passing. The flow is optimized for speed and token cost savings compared to full-model generation.

How this skill works

You write a contract-style prompt that defines data models, API routes, algorithm steps, and implementation rules. The skill calls the hosted LLM to generate project files and writes them directly to disk, returning metadata about files produced. Tests are run against the generated code and any failures are fixed with small edits using an edit tool, repeating until tests pass.

When to use it

• Generate algorithmic or infrastructure code (rate limiters, parsers, state machines).
• Create projects that require multiple files (3+ files) or heavy boilerplate.
• Scenarios with token-constrained sessions where cost matters.
• When you want fast first-pass generation and can tolerate 0–2 small fix cycles.

Best practices

• Write a precise contract prompt: explicit data models, API routes, algorithm steps, and rules.
• Include exact types and field names in data contracts and explicit route signatures in API contracts.
• Design tests that exercise edge cases so fixes remain small and focused.
• Keep fixes surgical: prefer targeted edits for 1–4 lines rather than broad rewrites.
• Choose models based on tradeoffs: default for speed/value, larger models for fewer fixes.

Example use cases

• Generate a FastAPI service with clear Pydantic models and endpoints, then run integration tests.
• Create multi-file libraries (parser + tokenizer + utils) where boilerplate dominates.
• Implement state machines or rate limiters with defined state serialization and reset endpoints.
• Scaffold complex algorithmic components that need consistent interfaces across files.
• Rapidly prototype service backends for tests where POST /reset clears state for repeatable runs.