reference guide

Train by Example — in ChatGPT, Claude, or Gemini

← Back to the exerciseCockpit

Doing the exercise in a chat tool

This guide replicates the Train by Example exercise in an ordinary chat product instead of the course's Pattern Trainer. The exercise is identical — the rule, the three parts, the capture. Two things change. First, a chat tool gives you no system box, so the one permitted instruction moves to the top of your message. Second, a chat conversation accumulates: every pattern you send becomes context for the next one, which quietly contaminates your experiments. The setup below handles both.

Set up your tool first

  1. 1
    Use a temporary chat.
    Temporary chats disable memory and personalization. In ChatGPT: click the model name → "Temporary chat". In Gemini: open the menu → "Temporary chat". In Claude: memory is off by default in a fresh chat; if your account has memory features enabled, turn them off in Settings.
  2. 2
    Start a NEW temporary chat every time you switch to a different pattern.
    This is the part people forget. In one continuing chat, your earlier patterns are still in the conversation — the model may blend them into its completions, and your experiment stops being clean. Iterating on the SAME pattern in one chat is fine; a new pattern gets a new chat.
  3. 3
    Put the one instruction at the top of every message.
    It replaces the trainer's system box: "Complete the pattern. Output only the completion, with no explanation." Everything below it must be examples only — no other sentences.
  4. 4
    You have no Save button — your course document is the library.
    Every time a pattern works, copy the whole message verbatim into the document. That library is what we read together at the end.

Part 1 · Classification

Start from this working pattern, then rebuild it for your own domain — three to five labels, four to six examples, ending with one left incomplete:

A complete Part 1 starter — paste into a fresh temporary chat

Then probe it exactly as the exercise says: delete examples one at a time to find how few still classify correctly; write a genuinely borderline case as the incomplete example; add one example that teaches the boundary. Each probe is an edit to the SAME message — re-paste the edited version into a fresh temporary chat and send.

Part 2 · Problem solving

Same move, with the procedure shapes. The skeletons, ready to fill in with a worked example from your domain:

The without-thinking skeleton

The with-thinking skeleton

Send each skeleton — filled in — from its own fresh temporary chat, give both the same new problem, and compare the completions side by side.

Part 3 · Declarations

No change from the main exercise: find a misbehaving pattern, repair it with the smallest possible declaration — a labels: {…} line, one dash rule, a range like 0..10 — added to your message above the examples, below the instruction line. Re-paste into a fresh temporary chat and confirm the fix.

The capture

The capture is identical to the main exercise — the misbehavior, the declaration with its word count, the corrected completion, and the two discussion questions. Copy everything into your course document as you go, since the chat tool saves nothing for you: Back to the exercise →