Skip to content
PrepMint

Guide series

What decides the answer

5 parts · 55 min in total.

Ask why an AI tool did something strange and almost every good answer turns out to be about the same thing: what text was actually in front of the model, and who decided. Which file it really read. What fell out of the conversation. What a search chose to return on your behalf. Most of what we publish is about that question, because most of the time it is the right one.

This series is about the other half — the part that is already decided before the meaning of your prompt is considered at all. Your words are broken into tokens the model never sees as letters. It already knows a version of the world with a date on it, and cannot tell that knowledge apart from what you gave it. An example you supplied outranks the rule you wrote next to it. The position of your last edit, not its size, sets what the turn costs. And an agent stops when it judges it is finished, which is not the same as being finished.

None of the five is a setting you can change. Each one is a mechanism you can predict once you know it is there, and each part below names the fingerprint it leaves — what the failure looks like from where you are sitting — so you can tell which of the five you are actually looking at.

Read them in order. Each part assumes the one before it, and every claim is checked against the vendor's own documentation with the sentence it rests on quoted and linked.

  1. Part 1: Why the model can't count the letters in a word

    Strawberry has two r's, your 100-word answer runs to 137, and the limerick does not rhyme. Those are not three mistakes. Text becomes tokens before the model sees it, and everything below the token — letters, characters, sounds — was discarded at the door.

    11 min read

  2. Part 2: Why it confidently used an API that no longer exists

    The code ran, and it was wrong. When a model answers about a library you did not paste, it is answering from weights with a date on them — and nothing in the output distinguishes what it knows from what it reconstructed. The four fingerprints of stale parametric knowledge, and the habit that fixes most of them.

    11 min read

  3. Part 3: It followed your example instead of your instruction

    You wrote the rule, then pasted an example under it to help — and the output matched the example. A demonstration is a stronger signal than a description, it shows everything about itself rather than the one property you meant, and nothing warns you when the two disagree.

    11 min read

  4. Part 4: Where you edit your prompt decides what it costs

    The session was fast for twenty turns, then you added one line to the top of a file and the next answer crawled. A prompt cache is keyed on an exact prefix, so cost and latency depend on where you changed something, not how much. The ordering rule that falls out of it.

    11 min read

  5. Part 5: Why the agent stopped before it finished

    You asked for five things and got one, with a confident summary and no error. An agent loop ends when the model writes a message containing no tool call — a judgement nothing checks against your task. The five ways a run ends early, each fingerprint, and the phrasing that fixes it.

    11 min read