claude-scaffold: Skill for Claude Code

NLP / SLM Patterns

When to Load This Skill

Load when working with: local language models, anonymization/PII detection, text classification, NER, Ollama, vLLM, Presidio, spaCy, Hugging Face Transformers, SLM inference pipelines.

Philosophy

Local models run locally during development, then deploy to YC GPU VM for production. The adapter interface is identical in both environments — only the endpoint URL changes via .env.

Anonymization projects deal with real personal data. Default posture: treat all input as sensitive until proven otherwise. Never log raw text containing PII.

Model Serving Options

Ollama Local Setup

import httpx
from src.project_name.core.config import settings


class OllamaAdapter:
    def __init__(self) -> None:
        self._base_url = settings.ollama_base_url
        self._model = settings.ollama_model

    async def generate(self, prompt: str, system: str = "") -> str:
        async with httpx.AsyncClient(timeout=120.0) as client:
            response = await client.post(
                f"{self._base_url}/api/generate",
                json={
                    "model": self._model,
                    "prompt": prompt,
                    "system": system,
                    "stream": False,
                },
            )
            response.raise_for_status()
            return response.json()["response"]

    async def chat(self, messages: list[dict]) -> str:
        async with httpx.AsyncClient(timeout=120.0) as client:
            response = await client.post(
                f"{self._base_url}/api/chat",
                json={"model": self._model, "messages": messages, "stream": False},
            )
            response.raise_for_status()
            return response.json()["message"]["content"]