How to Give ChatGPT Prompts That Are Easy to Check

The most common problem with a ChatGPT prompt is not that it is too short—it is that it is not checkable. The model returns an answer, and you have no clear way to tell whether that answer is right, complete or in the right format. A prompt structure that includes a brief, a source, a constraint and an acceptance check solves that problem. Tools like fenbs store tested prompts so you can reuse the ones that work without rewriting them each session.

You do not need elaborate prompt engineering for most tasks. The four-part structure here takes thirty seconds to apply and makes the output significantly easier to review.

The four parts of a checkable prompt

Most prompts that fail do so because they are missing one of these four elements. Not every prompt needs all four, but knowing what each one does helps you decide which to include.

• Brief: what you want, stated directly. Not "can you help me with" but "write", "summarize", "list" or "compare". One sentence.
• Source: what the answer should draw on. Pasting in the relevant text, naming a document, or specifying a field of knowledge. Without a source, the model draws on general training data, which may not match your situation.
• Constraint: what the answer should not include, or how it should be formatted. Length, format, audience, tone, topics to exclude. Constraints prevent the model from filling space with material you did not ask for.
• Acceptance check: how you will know the answer is correct or complete. A question you expect the model to answer, a format you can compare against a template, or a specific claim you will verify. This is the part most people leave out.

Worked example: summarizing a document

Say a legal assistant needs to summarize a lease agreement for a client. A vague prompt would be: "Summarize this lease." A prompt with the four-part structure looks like this:

Brief: Summarize the key terms of this lease agreement for a non-lawyer.
Source: [paste the relevant sections of the lease]
Constraint: Plain language, no legal jargon. Maximum 200 words. Cover rent amount, lease term, termination conditions, and maintenance responsibilities only.
Check: At the end, confirm which of the four topics you covered and flag any that were not addressed in the document.

The acceptance check at the end is explicit: the model should confirm which topics it covered. That gives you something to verify before you pass the summary to the client.

Worked example: drafting an email

A marketing manager needs to write a follow-up email after a missed deadline:

Brief: Draft a follow-up email to a client whose project deadline was missed by one week.
Source: The delay was caused by a supplier issue, now resolved. New delivery date is November 5.
Constraint: Professional and direct. One paragraph, under 100 words. Do not apologize more than once. Do not promise anything beyond the new delivery date.
Check: The email should include the new delivery date, a one-sentence explanation of the cause, and a single apology. Flag if any of these are missing.

Giving the model the right source material

The source part of a prompt is where most errors originate. If you ask ChatGPT to summarize a document without pasting it in, the model either refuses or summarizes from general knowledge, which may be wrong for your specific case. If you ask it to advise on a policy without including the policy text, it advises on what policies typically say, not on what yours says.

Paste the relevant section rather than the entire document when you can. A 20-page contract is harder to work with than the three clauses that matter for the question you are asking. The model's attention is finite; give it the material it needs and not much more.

What constraints are most useful

The most useful constraints are ones that prevent the model from filling space. Without them, responses tend to be longer and more hedged than necessary. These constraints reliably improve output:

• Word or sentence count: "Under 150 words" or "Three sentences maximum".
• Format: "Plain text, no bullet points" or "A numbered list with five items".
• Audience: "Write for a person with no technical background" or "Assume the reader is a practicing attorney".
• Exclusions: "Do not include recommendations, only observations" or "Do not mention competitor products".

When a prompt does not produce what you need

If a response misses the point, add a tighter constraint before trying again. "That was too long—rewrite it in two sentences" gives the model specific direction. "That's not what I meant" does not. Describe what is wrong with the output and ask for a targeted correction.

If a specific type of prompt consistently needs the same correction, that correction belongs in the prompt as a constraint. If every email draft comes back too formal, add "conversational, not formal" to the constraint section rather than correcting it after the fact each time.

How to do this with fenbs

1. Create a free account. Your personal board is a good starting point for a prompt library.
2. Add a Skill or prompt for each task type you repeat: summarize a document, draft a follow-up email, compare two options. Include a brief, a source requirement, constraints and an acceptance check in the template.
3. Add standing rules to Decisions and rules: which audience to write for, which topics to exclude from all responses, which format applies to client-facing work. Connected assistants read these rules first.
4. Connect your AI app via Integrations. Ask the assistant to list your saved prompts at the start of a session and use the relevant one for the task at hand.
5. When a prompt produces a consistently better result after a correction, update the saved version with that correction as a constraint.

See system prompt examples for ready-made instruction templates, and ChatGPT prompts for work for prompts organized by work task.