Prompt Engineering with Local LLMs

Anyone can type a question into a chatbot. Getting an AI to reliably produce work you’d actually put your name on, whether that’s a defensible finding, a tuned detection, or a configuration review that identifies the real escalation path, is a different but learnable skill. Building off the workshop Keeping Things Local: Build Private LLMs for Your Team, you’ll take the local LLM you built and learn to prompt it for reliable, consistent results with fewer hallucinations. Didn’t take the first workshop? No problem: a ready-to-go VM is provided.

This is a hands-on, four-hour lab. You’ll learn the two layers of prompting that most people blur into one: building a strong prompt before you hit enter, with the right role, context, constraints, and output shape, then running the refine-and-verify loop after the first draft comes back. You’ll practice on the jobs you actually do: writing findings, analyzing configs, drafting detections, and validating exploit preconditions in a lab. From there you’ll split planning from execution across sessions so unrelated context doesn’t pile up, and package a prompt worth reusing into a Skill your local model can load on demand. You’ll also spend time on the adversarial side, learning how prompt injection, jailbreaking, and data-extraction attacks work and why, so you can recognize them in the wild and craft and evaluate them during authorized security testing. Because nothing hurts your credibility faster than shipping a report with “AI wrote this” fingerprints on it, you’ll finish with a checklist for scrubbing AI slop before anything reaches a client.

Who Should Take This Workshop

Security practitioners who want to get real, defensible work out of an AI, and who’d rather keep sensitive prompts on hardware they control. It’s a strong fit for:

• Red teamers and penetration testers who want to speed up findings, tooling, and PoC validation without shipping client data to a third party
• Blue teamers and detection engineers who want to draft and tune Splunk and EDR detections faster
• Consultants and MSSP staff who write a high volume of client-facing deliverables and can’t afford AI fingerprints in a report
• Anyone who has used an AI chatbot, been underwhelmed by the results, and suspects better prompting would get better results

What You’ll Learn

By the end of the workshop, you will be able to:

• Pick the right model for the task, and explain the tradeoffs between proprietary, open-weight, uncensored community, and abliterated models
• Spend tokens deliberately: recognize context rot and the standing tax of repeatedly reloading context, and apply concrete techniques to cut waste
• Build strong prompts with the right role, context, constraints, and output shape, then run a disciplined refine-and-verify loop on real security tasks: findings, config analysis, detections,and proof-of-concept tooling
• Separate planning from execution across multiple sessions so a polluted context never derails a build
• Package repeatable prompt workflows into reusable Skills that extend your LLM’s capabilities
• Craft and recognize injection, jailbreaking, and data-extraction prompts, and understand why they work
• Scrub AI “tells” out of client-facing writing so your deliverables read like a human wrote them
• (Optional, building on the prior workshop) Reach your model from anywhere using a sandboxed, dedicated AI box
• Build lightweight evaluation sets so you can compare prompt revisions and identify real improvements