Day 2 — The anatomy of a prompt

Seven parts. Most bad prompts are missing five of them.

The seven parts

Part	Answers	Example
Role	Who should it be?	You are a senior data analyst
Task	What should it do?	Write a SQL query that...
Context	Why, and for whom?	For a weekly review read by the ops head
Data	What is it working with?	Table and column names, sample rows
Format	How should it come back?	A table with three columns
Constraints	What are the limits?	Under 100 words, no CTEs
Example	What does right look like?	One worked case

The score is a checklist, not a grade The workbench counts which parts are present. A prompt scoring 70 with exactly the right parts beats one padded to 100 with a role and an example it never needed. Use the missing-part list as a prompt to think, not a target to max out.

You will not always need all seven "Explain what #REF! means" needs no role, data or example. But when an answer comes back wrong or generic, the fix is almost always a missing part - and running down this list finds it faster than rewording at random.

Role sets the vocabulary

Telling it who to be changes the register, the assumed knowledge and the level of detail. It is one line and it costs nothing.

Task needs a verb

"Sales data" is not a task. "Summarise", "compare", "convert", "list", "debug" are tasks. If a prompt has no verb, the model has to guess what you want done, and it usually guesses "describe".

Add one verb and a target, and the score moves immediately:

Context is the part people skip

The model has no idea what your company does, who reads your reports or what tool you are stuck with. Two sentences of context prevent most of the useless answers.

Building one up, part by part

Start with the bare task, then add a part at a time and watch the score climb. Edit the box below - it re-scores as you type.

Suggested order to add them Task first, because without it nothing else helps. Then data, then format - those two remove the most guesswork. Role, context and constraints next. An example last, because it takes the most effort to write and you often will not need it.

The 80/20 of prompting

If you only ever remember two things: say exactly what you want back, and give it your actual data. Format and Data carry more weight than the other five parts combined, because they are the two the model genuinely cannot guess.

Try these yourself

Name the seven parts from memory.
Take "make a chart" and add a task verb and a format.
Write a prompt with all seven parts and score above 90.
Explain why context prevents useless answers.
Say which two parts matter most, and why.

← PreviousDay 1 — What a language model actually is Next →Day 3 — Being specific - the biggest lever