Prompt Engineering Guide

You don't need to memorise tricks to get better answers from an LLM. You need a repeatable structure. This guide shows the one TPEE uses — twelve sections, one quality score — with a worked example you can copy.

By Jacques Botte, founder of Toptronic^®. Last updated 12 September 2026.

Why free-typed prompts fail

Most people type a paragraph into ChatGPT and hope. The model is doing its best with an incomplete brief, so it guesses — and a guess about what you wanted is a wasted answer. The three failure modes are the same everywhere:

Missing context. The model doesn't know your situation, so it fills the gap with something generic.
Vague instructions. "Make it good" gives the model nothing to optimise against.
No examples. You know what a good answer looks like; the model doesn't, unless you show it.

None of this is the model being dumb. It's the brief being incomplete. A structured prompt fixes all three at once.

The 12-section method

TPEE splits every prompt into twelve sections. You don't have to fill all of them every time — a short question may only need Persona, Instructions and Tone — but each one you fill removes another way for the model to guess wrong:

Persona — who the model is (a senior Rust developer, a patient teacher).
Instructions — the task, stated plainly.
Context — the background that makes the task make sense.
Technical Specifications — hard constraints (language version, limits, formats).
Format — how the answer should be laid out (Markdown, JSON, bullet list).
Audience — who will read the answer, so the model pitches it right.
Tone — formal, friendly, terse.
Data — the raw facts, references or source material.
Examples — one or two examples of a good answer.
Graphics — whether you want diagrams, tables or charts.
MCP Servers — external tools the model may call.
Agents — sub-agents or roles for multi-step work.

A worked example

Before (what people actually type):

"write me a python script that reads a csv and does some analysis"

After (the same request, structured):

Persona: You are a senior Python data engineer.
Instructions: Write a Python script that reads a CSV file and reports summary statistics for each numeric column.
Context: The file is exported weekly from our sales system and may contain empty cells.
Technical Specifications: Python 3.11, standard library only, no third-party packages.
Format: Output as a Markdown table of column name, count, mean, min, max.
Audience: A non-technical manager who will read the summary.
Tone: Concise, no jargon in the final table.
Examples: If a column has missing values, show the count of missing values in parentheses.

The structured version will produce a correct, complete script on the first try. The free-typed version will probably need two or three follow-ups — each one costing tokens and time. That difference is the whole point.

Why structure saves money

Every time you re-run a vague prompt because the answer missed the mark, you pay again. On a large-context model a single retry can cost hundreds of thousands of tokens. Shortening and clarifying the prompt before you send it is the cheapest optimisation there is — and the 12-section structure plus a 0–100 score does that for you. TPEE itself never calls a model; the savings are in how you prompt.

TPEE automates this checklist — from $20, one-time, no subscription.