The one-shot prompt
A vague prompt costs five rounds of "no, not like that." A complete one gets the right result the first time. This is the checklist, with the research behind it.
1 minute 53 seconds · animated infographic · sound on
Why this matters
Whatever you leave out of the prompt, the model invents: the goal, the audience, the definition of done. One-shot prompting is just leaving no gaps worth guessing about. It is not a talent. It is a habit with a checklist.
The four-part anatomy
- Goal. What you want, and where. Which file, which page, which account. "Improve the email" is a wish; "shorten the follow-up email to under 100 words" is a goal.
- Context. The facts the model cannot guess: the exact error message, who the audience is, what was already tried, the constraint that matters.
- Output spec. The shape of done: format, length, tone, what to include and what to leave out. If you want a table, say table.
- Done check. How you will know it worked: a test you can actually run. "Done when the checkout page loads without the error."
Miss one part and the model fills that gap for you, confidently. The why-AI-guesses explainer covers what happens next.
The context rules
These come from research, not taste (both studies are on the AI safety hub):
- Only what is needed. Model accuracy falls as input grows, well before the context window is full. Every extra paragraph you paste is a place for your one important fact to get lost. Trim before you send.
- Edges beat the middle. Models recall the start and end of a prompt best and lose the middle. Put the goal first and the hard constraint last, never buried between pasted logs.
- One job per prompt. Three asks in one message means the model averages them into one mediocre answer. Split them into three prompts, or delegate them as three jobs.
The worked example
The F: fix it. No target, no context, no way to tell when the job is done. This earns the five rounds it causes.
The A: Fix the checkout bug on cart.js: clicking Pay throws TypeError line 41. Done when the checkout page loads without the error.
Same request. The A version has all four parts in two sentences: goal (fix the checkout bug on cart.js), context (clicking Pay throws TypeError line 41), and a done check you can run. A one-shot prompt does not need to be long. It needs to be complete.
Verify before you send
The last habit: grade the prompt, not just the answer. In the Cronus composer, the A± button grades your draft A to F against exactly this anatomy: clarity, context, output spec, structure, examples, constraints. Each grade tells you which part is missing in plain words.
Fix it for me then writes the A-standard rewrite, and you choose: use it, or keep yours. It never replaces your words on its own. The composer guide shows the whole flow on screen.
The checklist, to keep
- Goal: what and where?
- Context: what can the model not guess?
- Output spec: what shape is done?
- Done check: what test proves it?
- Trimmed: is anything in here the model does not need?
- Edges: goal first, hard constraint last?
- One job: am I asking for exactly one thing?
- Graded: did the rater give it an A?