AI engineering · Revise in 10 minutes

Cheat sheet: AI engineering

Every module on one page: its memory hook and the points worth saying out loud. Print it (Ctrl/⌘ P) and re-read it before a practice session; open the page behind any card when something doesn't come back to you. Built automatically from the pages, so it always matches them.

Level 1 · How AI works

Tokens and prediction

“A very well-read autocomplete”

It predicts the next token, adds it, and predicts again. Brilliant at plausible, not a database of truth.

Avoid

Trusting specific facts without a source
Counting words instead of tokens
Judging from one run

The context window

“The desk, not the library”

The model works only from what's on its desk right now. The desk is a fixed size, clutter makes it worse, and it's cleared between sessions.

Avoid

One endless session
Pasting whole files and full logs
Connecting every MCP server "just in case"

Training and hallucination

“Closed book guesses, open book cites”

From memory alone, a model fills gaps with plausible guesses. Give it the source and ask it to cite, and you can check every claim.

Avoid

Asking for citations and not opening them
Asking "are you sure?"
Using a model for recent facts without search

Reasoning, tiers and cost

“Three dials on the bill: tier, effort, tokens”

Every quality, speed and cost decision is one of three dials. Start low and raise one dial at a time, only when the result isn't good enough.

Avoid

Maximum effort on everything
Forgetting thinking tokens in estimates
Comparing vendors' price lists directly

Level 2 · Working with AI

Prompting and context

“Brief it like a brilliant new hire”

Assume skill, not context. Give the goal, the context, the constraints with reasons, an example, the format and what "done" means.

Avoid

No definition of done
Rules without reasons
Shouting

Tools and function calling

“The model writes the order; your code cooks”

The model only ever produces a precise request. Your code decides whether to run it, runs it, and brings back the result.

Avoid

Authorisation in the prompt instead of the tool
Vague names and descriptions
Dumping huge results

Agents: the loop

“Act, look, decide: a loop on a leash”

An agent repeats three moves until the goal is met. The leash (a definition of done, budgets, permissions and you) decides how far it may run.

Avoid

Full autonomy on your own machine
No definition of done
Using an agent for a fixed process

The agentic coding workflow

“Map, plan, slice, prove”

Understand before changing, agree before building, change in small steps, and never merge without evidence.

Avoid

Merging on the agent's word
Skipping the spec
Giant diffs

Level 3 · Configure your tools

Instructions and memory

“Leave a note on the desk”

The agent forgets everything overnight. AGENTS.md is the short note that's on the desk every morning: commands, layout, rules with reasons, and the traps you've hit.

Avoid

Secrets in instruction or memory files
Instructions that grow forever
Rules without reasons

MCP

“USB-C for AI tools”

One standard plug. Build or install a server once, and it works in every tool that speaks MCP. Like any device you plug in, check what it can do before trusting it.

Avoid

Production write access
Secrets written into config files
Connecting everything

Agent Skills

“Recipe cards in a drawer”

Every skill's title is visible; the full recipe comes out only when that dish is ordered. Write the title so the right card gets pulled, and the recipe so it ends with a taste test.

Avoid

Running unreviewed skills
Vague descriptions
A wall of prose

Subagents, hooks, plugins

“Specialists think, guards enforce, kits share”

Subagents take on focused work in their own context. Hooks enforce rules no prompt can bend. Plugins hand the whole setup to your team in one install.

Avoid

Installing plugins without reading them
Loose hook patterns
Subagents with too many tools

Automation and CI

“Every robot gets its own key card”

Headless agents run with no one watching, so each job gets only the doors it needs, and a person keeps the key to main.

Avoid

Write tokens on jobs that read untrusted text
Letting automation merge or deploy
No budget or timeout

Level 4 · Build AI features

Model APIs and structured output

“Ask for a form, not a letter”

When your code needs data, give the model a schema to fill in, then check the form before you file it.

Avoid

API keys in client-side code or repositories
Ignoring the stop reason
Trusting structured output for business rules

Embeddings and RAG

“Index, fetch, cite”

Prepare the library once, fetch the few right pages for each question, and make every answer point to its page.

Avoid

Retrieving what the user isn't allowed to see
Embeddings only
Chunks without context

Building agents and MCP servers

“Brain, hands, loop, leash”

Every agent is a model that thinks, tools that act, a loop that repeats, and limits that keep it safe. Leave out the leash and it isn't ready to ship.

Avoid