// the field manual
How to build with AI that survives production
A working curriculum for building real software with AI, grouped into modules. Start with the foundations, how these models actually behave, then work through spec, verify, guardrails, and the rest. Pick a module to see its lessons.
- // module 00
How these things actually work
How models actually generate, drift, hallucinate, and forget. The floor under every other track.
8 lessons
Autoregression: the output is the thinking · Temperature and drift · In-context learning · +5 more
- // module 01
Spec before you generate
A prompt is a wish; a spec is a contract. The discipline that separates production from slop.
10 lessons
Spec before you generate · Why a prompt is not a spec · The six things every spec pins · +7 more
- // module 02
Start simple
Complexity is a loan. Only take it when the simple version has actually failed you.
9 lessons
Start simple · Complexity is a loan · When do you actually need an agent? · +6 more
- // module 03
Verify, don't trust
Generation got cheap. The review is the whole job now.
9 lessons
Verify, don't trust · Verification is the job now · Counterexample hunting · +6 more
- // module 04
Make the bad state impossible
If it must never happen, don't ask the model to be careful. Build it so it can't.
9 lessons
Make the bad state impossible · Don't ask, enforce · The 99% problem · +6 more
- // module 05
Reversible by default, gate the irreversible
Speed on what you can undo. A human on the one button you can't take back.
9 lessons
Reversible by default, gate the irreversible · Sort every action by whether you can undo it · Speed on the reversible, ceremony on the irreversible · +6 more
- // module 06
When it breaks, triage before you rebuild
The rewrite feels faster and almost never is. Diagnose the one line first.
9 lessons
When it breaks, triage before you rebuild · The rewrite reflex, and why regeneration feels like progress · Diagnose the one line before you throw away the whole thing · +6 more
- // module 07
Externalize your memory
The context window fills and forgets. Own your state; the thread is disposable.
9 lessons · gated finale
Externalize your memory · The context window fills, and then it forgets · Files beat ephemeral chat memory · +6 more
- // module 08
The edge
The advanced wing. Recursion, compound hallucination, and why a model's own output can start to feel like a mind, explained by the mechanism instead of the mysticism.
7 lessons
What recursion does to a language model · The what-next spiral, a case study in compound hallucination · Why recursive output starts to feel like a mind · +4 more
// common questions
- What is vibe coding?
- Vibe coding is building software by directing AI to write the code instead of typing it yourself. You describe what you want, the model generates it, and your job becomes steering and verifying rather than authoring every line.
- Do I need to know how to code to build software with AI?
- No, but you do need to think like a builder. You do not have to write the syntax, but you have to specify what you want clearly, verify what comes back, and recognize when something is wrong. This manual teaches exactly that.
- What is a spec, and why does it matter?
- A spec is a short contract that says what the software must do, what must always be true, and what counts as broken, written before the AI generates anything. It is the single biggest thing separating production software from AI slop.
- Why does AI make things up?
- A model predicts the most likely next words and has no built-in signal for when it does not know something, so it fills gaps with confident guesses. That is the default behavior, not a glitch, which is why you verify instead of trust.
- How do I verify what AI generates?
- You check the output against something outside the model: a spec, a test on real values, a second model, or your own read. Verification is now the main job, because generating code got cheap and being sure it is correct did not.
- How is this different from just chatting with an AI assistant?
- Chatting gets you a quick answer. Building software that survives real users needs discipline the chat does not give you: specs, verification, guardrails, and reversibility. This manual is that discipline.
- How do I keep AI from forgetting what we are working on?
- The context window is finite and degrades as it fills, so you externalize your memory into files you control instead of trusting the conversation. The thread is disposable; your saved state is what persists.