---
title: AI Steering
emoji: 🧭
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 5.9.1
app_file: app.py
pinned: false
license: mit
---

# Core AI Steering Skills

Interactive, **fully offline** demo of four practical techniques for steering
LLM applications — no API key required, no live model calls:

- **Reasoning Enhancement** — Chain of Thought and Tree of Thoughts prompting
- **Context Window Management** — conversation buffers that condense instead of forget
- **Semantic Guardrails** — keyword prefilter (real code) + semantic classification
- **Model Routing** — a cheap classifier routes requests to the matching model tier
- **Output Guardrails** — regex PII redaction of model responses (real code)
- **Cost Tracking** — per-tier pricing showing what routing saves vs. all-Opus (real code)

This Space runs in **simulated mode**: local Python heuristics stand in for
the Claude API calls the real `ai_steering` package makes, so you can explore
the mechanics of each skill for free. The guardrail keyword prefilter, PII
redaction, and cost tracker are the actual production code, since they're
designed to never need a model call.

Source: https://github.com/aabhimittal/ai-steering — the GitHub package is
fully Claude/Anthropic-backed; set `ANTHROPIC_API_KEY` and run `examples/*.py`
locally to see live model output.