GenAI Field Guide

Start here

GenAI Field Guide

This site helps people learn what generative AI is and how to use AI tools well: where your text may go, which tool fits which job, what projects and agents actually are, and the checks to run before you rely on an answer. One topic per page, a few minutes each.

1 Getting started

What it can do, where your text may go, which model and tool, and the habits that make it safe.

What it can actually doSix kinds of work it is good at, what to watch for in each, and ten things to try this week.Where your text goesWhat you may put in which tool, and why IL5 clears security, not correctness.The lineup on GenAI.milWho is on GenAI.mil, who isn't, and the four ways vendors differ.Choosing a modelWhy Grok over ChatGPT, why an older version over a newer one, and when to turn thinking on.Which tool for which jobA one-table cheat sheet.Prompting that survives contactFour things to include, one thing to ask at the end.ProjectsThe container that remembers. SOPs and style guides go here.Documents, images, and videoAttach it, project it, or retrieve it.

2 Under the hood (just enough)

Just enough about how the tool works to use it well: tokens, spec sheets, thinking modes, the prompt, the window.

What an LLM is (and isn't)A next-token predictor: fluent, confident, memoryless.TokensThe unit everything is measured in.Reading a model's spec sheetWhat 35B, 128k, Q4, "open-weight," and a knowledge cutoff mean, in one page.Thinking modesWhat "reasoning" or "thinking" does, when to turn it up, and when it just makes you wait.What's actually in the promptYou typed one line. The model received thousands.The context windowWhy 1M tokens is never 1M tokens of your chat.CompactionLong chats get summarized. Detail is lost.The harness is the productWhere to look when a tool behaves differently.

3 Where it fails

Common pitfalls, hallucination, verification, and the pattern that keeps you safe.

Common pitfallsThe twelve mistakes people actually make, and the one-line fix for each.HallucinationWhy it guesses, what helps, what only changes the tone.The verification checklistFive questions before anything leaves your desk.Not fire and forgetDraft, check with code, fix, gate, human.Prompt injectionHidden instructions in a document are still instructions.

4 Agents and beyond

What agents, tools, subagents, RAG, and the API actually are.

AgentsAn LLM in a loop with tools, until a stop condition.Tools, connectors (MCP), and skillsHow a model gets abilities.Subagents and sessionsFresh helpers that return only a summary.Coding sessions: where the output goesOptional, for people who run coding agents. Where program output goes, and why errors get forgotten.RAG, retrieval-augmented generationAnswer about your documents without retraining.API vs. the web interfaceSame model, full control, no rails.

5 Reference

Myths, glossary, training, and the evidence.

Myths, bustedEvery myth in the guide, with the one-line fix.GlossaryOne line each, grouped by topic.Training and certificationFree first, then low-cost.The evidenceThe papers behind the claims.About this guideWhat this is and how to use it as a course.

Five habits

Start with these five habits. Before anything leaves your desk:

  1. Did I give it the source, or is it answering from memory?
  2. Did I tell it that "I don't know" is acceptable?
  3. Every number, date, name, and citation: checked against what?
  4. Did I ask a second, fresh chat to find what is unsupported?
  5. Would I sign this without the AI's name on it? If not, it is not done.

About this guide

Written for people who use these tools at work, including on GenAI.mil, and for anyone new to them. Platform and vendor facts were checked on 31 August 2026 and are dated on the pages that carry them; the mental models do not change. Nothing here is official guidance from any organization; your own organization's rules and the platform itself take precedence. See About for how to use this as a course, and The evidence for the papers behind the claims.