Get the AI Field Guide in your inbox
AI Field Guide
Short, practical entries on getting the most out of AI tools. Browse by category, or search from the sidebar.
Prompting
Critique loop
A critique loop has one model do the work and a second model review it before you accept it. A simple way to raise quality without more prompting.
Having it ask you questions
One extra line makes an AI interview you before it starts, so it surfaces what you forgot to mention instead of guessing at it.
Voice input
Talking instead of typing gets far more context into a prompt. People speak around 150 words a minute, and type a fraction of that.
Writing good prompts
Most weak answers come from thin prompts. Role, context, task, and format turn a one-line question into something a model can act on.
Concepts
Agents
An AI agent takes a goal, picks its own tools, does the work, and checks the result. What that means in practice, and where agents are used today.
Context engineering
Context engineering is deciding what a model sees on each call. Why long chats get worse, and the habits that keep an agent sharp.
Effort level
Effort level is a dial for how hard a model thinks before it answers. Turn it up for hard problems, turn it down for speed and lower cost.
Evals
An eval is a test for AI output: run a prompt or skill over real examples and score the results. Which checks to automate, and which need a judge.
Fine-tuning
Fine-tuning trains an existing model further on your own examples, so the behavior you want is baked in instead of prompted every time.
Hallucinations
When a model doesn't know something it fills the gap with a confident guess: invented statistics, fake citations, quotes nobody said. How to catch it.
Harness
A harness is everything wrapped around a model that turns it into a working [agent](/ai-field-guide/agents). Tools, memory, workspace, the loop that keeps it running.
Local vs cloud AI
AI in the browser runs in the cloud, outside your computer's trust boundary. Desktop AI tools are installed locally, so they can reach your files.
MCP
MCP is a standard way to plug an AI into your tools and data, replacing a custom connector per service with one shared protocol.
Model routing
Model routing sends each task to the cheapest model that can still do it, instead of defaulting to your most expensive model for everything.
Open, open-weight, and proprietary models
Proprietary models stay on the company's servers. Open-weight models you download and run yourself. What each label actually means.
Personal Agent Protocol (Poppy)
Personal Agent Protocol (Poppy) is a draft standard for how your AI agent signs in and acts with a company for you, and how it relates to MCP.
Prompt injection
Prompt injection hides instructions inside text an AI reads, and the model follows them as if they came from you. Why it works, and how to limit it.
RAG
RAG searches your documents for the relevant parts, then hands those to the model with the question, so the answer comes from your data.
Skills
A skill is a folder that teaches an agent to do one task your way, every time. What goes in a SKILL.md, and how to keep a library of skills useful.
Tokens
A token is the chunk of text a model reads and writes in, roughly four characters. Why pricing and context limits are both counted in them.
Vibe coding
Vibe coding means describing what you want in plain language and letting AI write the code without reading it. Where it works, and where it breaks.
Models
Choosing an AI model
Claude, GPT, and Gemini all handle most jobs well. How to pick one, when switching is worth it, and why chasing the benchmark leader wastes time.
Claude
Claude is Anthropic's family of AI models, reached through the Claude apps, the API, or Claude Code. What it is known for and how people use it.
DeepSeek
DeepSeek is a family of open-weight models released under an MIT license, known for matching pricier closed models on reasoning and coding.
Gemini
Gemini is Google's family of AI models, known for very large context windows and strong handling of long documents, images, and audio.
GPT
GPT is OpenAI's family of models and the engine behind ChatGPT. What separates the two, and how people reach each of them.
Grok
Grok is the model family from xAI, used through the Grok app and inside X, where it can draw on real-time posts.
Hermes
Hermes is a family of open-weight models from Nous Research, fine-tuned from open bases like Llama to follow instructions closely.
Kimi
Kimi is Moonshot AI's open-weight model family, built for agent work and competing with the closed frontier on coding and reasoning.
Llama
Llama is Meta's family of open-weight models, and the one that did most to make open models mainstream. Download it and run it yourself.
Qwen
Qwen is Alibaba's large family of open-weight models, with many sizes and versions specialized for coding and vision, mostly on Hugging Face.
Skills
Design systems
AI coding tools are weak at design by default. Pointing one at a design system gives it proven components instead of inventing every button.
Karpathy coding guidelines
Coding guidelines drawn from Andrej Karpathy's take on where AI agents go wrong. Think first, keep it minimal, make surgical edits.
Progressive disclosure
Progressive disclosure keeps a skill's full instructions out of every AI call. Give the agent a little up front and let it reach for the rest.
Remotion
Remotion renders real MP4 videos from React code, so a coding agent can build an animated video from a plain description.
Taste Skill
Taste Skill is a set of design rules you drop into a coding agent so it stops shipping generic frontends. How to install it and what it changes.
Tools
Claude Code commands
/compact/compact summarizes a long Claude Code session so you can keep going without starting over, and you can tell it which details to keep.
/goalType /goal and describe what done looks like. Claude Code or Codex then keeps working and checks itself until it's finished, even overnight.
/insights/insights builds a report on how you've used Claude Code over 30 days: what you worked on, what slowed you down, and what to try next.
/loop/loop hands an agent a repeating job. It works in rounds, prompting itself each tick, until the work is finished or you stop it.
/remote-control/remote-control puts a live window into your running Claude Code session on your phone, so you can check progress and approve steps anywhere.
/rewind/rewind rolls your code and your conversation back to an earlier point, like a save point, so you can leave a bad path without losing everything.