viliawang-pm

Plan and optimize token budget allocation for LLM context windows. Use this skill when designing system prompts, RAG pipelines, or multi-turn agents that must fit within a model's context limit. Produces a token budget breakdown, identifies waste, and suggests compression strategies to maximize the useful information density within the available context window.

Design, build, and optimize Retrieval-Augmented Generation (RAG) pipelines from scratch. Use this skill when building a RAG system, diagnosing retrieval quality issues, choosing embedding models, designing chunking strategies, or optimizing the end-to-end pipeline for accuracy and latency. Covers naive RAG, advanced RAG, and modular RAG architectures.

Design and implement safety guardrails for AI agent systems. Use this skill when building production agents that need protection against prompt injection, jailbreaks, data leakage, uncontrolled tool use, and other adversarial attacks. Includes red-team testing checklists, defense-in-depth architectures, and monitoring strategies.

A thinking partner for product managers building Demos, Websites, or Apps. Use this skill when a PM shares a product idea or begins prototyping. Instead of judging ideas, it helps PMs articulate the intuition behind their thinking, sharpen problem definition, understand market context, envision vivid usage moments, and find the fastest path to validation — through guided conversation, not interrogation.

Build systematic evaluation frameworks for LLM applications and AI agent systems. Use this skill when you need to measure LLM output quality, compare model performance, evaluate RAG accuracy, benchmark agent reliability, or set up continuous evaluation pipelines. Covers both automated metrics and LLM-as-a-Judge approaches with bias mitigation.

Evaluate and optimize LLM prompts for quality, clarity, and effectiveness. Use this skill when reviewing system prompts, user prompt templates, or any instructional text destined for an LLM. Produces a scored rubric with actionable rewrite suggestions. Supports single prompts, A/B comparison, and batch evaluation workflows.