A clever prompt is written to sound impressive. A working prompt is written to be executed. Claude reads the second kind the way a contractor reads a work order: it looks for the job, the materials, the limits, and the shape of the finished thing. When one is missing, it fills the gap with a reasonable guess.
Why structure beats cleverness
Novelty does not survive a second run. A wordplay-heavy prompt can win a demo, but when the output is wrong it gives you nothing to edit. A structured brief is auditable: one task verb, named context, testable constraints, a format you can parse. It also travels from chat to Claude Code to a system prompt, because it never depends on a clever phrase landing.
The six parts of a working Claude prompt
Add them in this order. Each one narrows the next.
- 1Role. One line that sets the lens: staff engineer reviewing a diff, editor writing for a skeptical CFO. Skip the biography and the flattery.
- 2Task. One imperative verb and the outcome: review, extract, compare, plan. Two verbs usually produce half of each.
- 3Context. The source material the answer has to rest on: the excerpt, the diff, the error. Label it so the model knows what it is reading.
- 4Constraints. Rules written as tests: no invented numbers, keep the API stable, flag assumptions. Put the reader and a length cap here too.
- 5Output format. The shape you will use: five bullets, a table, JSON with named keys, a 40-word summary. If you must parse it, name the fields.
- 6Example. One short input-to-output pair. It locks a tone faster than rules do. Keep it correct; a bad example teaches the wrong pattern.
The same request, twice
You are a world-class genius editor. Make this landing page copy incredible and really make it sing. Think outside the box!
Role: conversion copywriter for a B2B analytics product. Task: rewrite the hero section below. Context: Audience is data leads at mid-size SaaS companies. Source: current copy, pasted below. Constraints: Keep the product name. No superlatives. Under 60 words. One call to action. Output: headline, 20-word subhead, three bullets. Example: "See every query. Cut the slow ones."
The second brief names the reader, bounds the answer, and gives the model a target to check itself against.
What vagueness costs
A vague prompt is not free. It is a prompt you will run again.
Take a 700-token brief that needs three attempts before the answer is usable. You paid for 2,800 input tokens to get one result, plus two answers you discarded. The retries usually happen because the answer was the wrong length, the wrong format, or written for the wrong reader. All three are fields you left out, not limits the model hit.
- Retries. Each attempt resends the entire prompt at full input price.
- Long answers. Loose scope produces padding, and every extra paragraph is billed as output.
- Review time. Hand-editing an almost-right answer never appears on an invoice, but it is paid in hours.
The cheapest prompt is the one that works on the first turn, not the shortest one.
Assemble, then test
- 1Write the task as one sentence. If it needs a second, split it.
- 2Add the smallest context that supports a correct answer, labeled.
- 3Write two to four constraints, including reader and length.
- 4Name the output shape and the fields you need back.
- 5Add an example only if the format or tone is unusual.
- 6Delete any sentence that would not change the output.
- 7Run it on ten real inputs, then fix the field behind each failure.
Structure is not bureaucracy. It is the difference between asking for an answer and specifying one. Claude will meet the level of detail you bring, and the six parts are the cheapest way to raise it.
Frequently asked questions
Does a role actually improve Claude output?
Only when it changes vocabulary, standards, or assumptions. A role that adds a costume but no judgment is tokens you can delete.
Can a prompt be too structured?
Yes. Every field must earn its tokens. If a constraint or example does not change the answer on real inputs, cut it.
Should every prompt include an example?
No. Examples are best for unusual formats and tones. For routine work they add input tokens without changing the result.
All content is original. The six-part structure was informed by Anthropic's public prompting guidance and the field patterns used in the TokenCalc AI Prompt Builder; no third-party course text is reproduced. No prices are hard-coded in prose; cost examples are computed at build time from the live catalog.