This is a working tool, not an essay. Keep it open while you write prompts, and run through it before you hit send. Each item is a question to answer yes to, with a one-line reason it earns its place. Skip items that genuinely do not apply to your task, but know why you are skipping them.
The checklist is organized into three passes: before you write, while you write, and before you trust the output. Most failed prompts skip the first and last passes entirely. The whole thing takes under a minute once you are familiar with it, and it catches the predictable mistakes that waste the most time.
We have kept it deliberately short. A checklist you actually run beats an exhaustive one you ignore. If you want the reasoning behind any item in depth, our best practices guide covers the why.
Pass 1: Before You Write the Prompt
These take ten seconds and prevent the most common failures.
- Have I defined what success looks like? You cannot judge output you never specified. Decide the bar before you write.
- Is this one task or several? Multiple tasks crammed together degrade quality. Split them or sequence them.
- Do I know who the output is for? Audience drives tone, depth, and vocabulary. "General audience" is a non-answer.
- Is this worth engineering, or a quick one-off? Reserve real effort for tasks you repeat; do not over-engineer trivia.
Why this pass matters most
The biggest determinant of output quality is decided before you type a word: knowing what you want, for whom, in what form. Our how-to guide builds these questions into its opening steps.
Pass 2: While You Write the Prompt
The construction checks. These shape the prompt itself.
- Did I lead with a clear imperative task? One command verb, stated directly, beats a polite ramble.
- Did I provide the context the model cannot know? Background, facts, source material. Assume zero prior knowledge.
- Did I separate data from instructions with delimiters? Tags or quotes prevent the model from confusing the two.
- Did I specify the exact output format, ideally with an example? Showing the format beats describing it.
- Did I add guardrails against fabrication? "Use only the provided text; if unsure, say so" stops invented facts.
- Is my most important instruction at the start or end? Models weight the edges more than the middle.
That is the core of the build. If you answer yes to all six, your prompt is already in the top tier of what most people send. The common mistakes guide explains the cost of skipping any one of them.
Pass 3: Before You Trust the Output
The verification pass. This is the one people skip, and the one that protects your credibility.
- Did I read the output critically against my success bar? Not skimmed, read. Find the biggest gap.
- Did I verify every fact, number, quote, and citation? Never trust facts the model produced from memory. Check them or remove them.
- Did I test on more than one input? A prompt that works once may break on the next case. Reliability across inputs is the real test.
- If I will reuse this, did I save it as a template? Parameterize the specifics and store it so the effort compounds.
The fabrication check is non-negotiable
Fluent, confident text reads as trustworthy even when invented. This single check, verifying facts before you ship them, prevents the most damaging category of AI error. Our case study shows a team that eliminated fabrication complaints precisely by grounding output and verifying it.
A Worked Example: Running the Checklist on One Prompt
Watch the three passes applied to a real task: drafting a product announcement email. In Pass 1, you define success ("a 120-word email a customer would actually read, announcing the new export feature, no invented capabilities"), confirm it is a single task, name the audience (existing paying customers), and decide it is worth engineering because you send these monthly.
In Pass 2, you lead with the imperative ("Draft an announcement email"), paste the real feature spec inside <spec> tags, specify the format (subject line, two short paragraphs, one call to action), and add the guardrail "describe only features in the spec; invent nothing." You place the no-invention rule at the end for weight.
In Pass 3, you read the draft against your success sentence, check that every claimed capability appears in the spec, run the prompt against a second feature to confirm it generalizes, and save it as a template with a slot for the spec. The whole sequence takes a few minutes and the result ships without a rewrite. That is the checklist doing its job.
Common Reasons People Skip the Checklist
Three excuses recur, and each has a rebuttal worth internalizing.
- "It slows me down." It saves time overall by preventing the rewrite loop that vague prompts create. The minute spent up front is recovered many times over.
- "I'll just fix the output." Fixing fabricated facts requires noticing them first, and confident wrong text is exactly what you fail to notice. Prevention beats correction here.
- "It's a simple task." Simple tasks need only the core checks, which take seconds. The checklist scales down; it does not demand full ceremony for a quick question.
How to Use This Checklist Day to Day
For high-stakes or reused prompts, run all three passes deliberately. For quick personal questions, Pass 2's "clear task" and "needed context" items alone will cover you. The skill is calibrating effort to stakes: a throwaway question does not need the full ceremony, but anything you will publish, send, or rely on does.
Over time, most of these checks become automatic and you will run them without thinking. Until then, keeping the list visible while you work is the fastest way to build the habits. The framework guide offers a mental model that ties these checks together into a single repeatable process.
Frequently Asked Questions
Do I have to run all three passes every time?
No. Calibrate to the stakes. A casual personal question needs only the core construction checks. Anything you will publish, send to a client, or act on deserves all three passes, especially the verification pass that catches fabricated facts.
Which single checklist item prevents the most damage?
Verifying every fact, number, quote, and citation before trusting the output. Fabricated information that looks confident is the costliest AI error because it ends up in work you ship. This check is the one to never skip on anything that matters.
What does "define success" actually look like in practice?
Write one sentence: "This prompt succeeds if the output is ___." For example, "a summary under 150 words with no invented facts and a friendly tone." Concrete criteria let you judge the result objectively instead of reacting with a vague sense of dissatisfaction.
Why test on more than one input?
A prompt can succeed on the example you happened to try and fail on the next one, because that first input did not stress the weak parts. Reliability across several varied inputs is the real measure of whether a prompt is ready, especially for one you will reuse.
How is this checklist different for 2026 specifically?
The fundamentals here are durable, but as models improve, the verification and grounding items matter more, not less, because fluent output is increasingly convincing. The checklist emphasizes guardrails against fabrication precisely because better models make confident errors harder to spot.
Key Takeaways
- Run three passes: before writing, while writing, and before trusting output.
- Define success and decide single-versus-multiple task before you type anything.
- Build with a clear imperative, full context, delimiters, exact format, and guardrails.
- Always verify facts, numbers, and citations before relying on output.
- Calibrate effort to stakes, and save reused prompts as parameterized templates.