Help me turn a recurring AI task into a reliable prompt with named variables. Task: [RECURRING TASK] Primary user of the output: [AUDIENCE] Inputs I have each time: [AVAILABLE INPUTS] Required output: [OUTPUT SPECIFICATION]
Examples of a good and bad result: [QUALITY EXAMPLES] First, identify the minimum 2-6 variables needed to produce the result consistently. Use uppercase square-bracket names, such as [AUDIENCE] or [SOURCE NOTES]. For each variable, state what belongs in it, what format it should use, and whether it is required or optional. Do not create variables that duplicate one another. Then write one copy-ready master prompt in my voice. It must: 1. list the variables at the top with short entry instructions; 2. distinguish source facts from assumptions; 3. specify the output format, sections, length, and audience; 4. tell the model what to do when an optional variable is blank; 5. prohibit invention of facts, figures, quotations, names, or policies; and 6. include a compact quality-control checklist before the final output. After the master prompt, provide a filled example using plausible sample values and a one-line version-control note describing what should change if the task changes. Self-check before answering: confirm every required result in [OUTPUT SPECIFICATION] has an input source or an explicit instruction. Flag any requirement that cannot be fulfilled from the listed inputs. Ask up to 3 clarifying questions only if a required input is missing.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [RECURRING TASK] | Describe the repeatable task you want the AI to perform. | Draft a weekly customer-success account update from call notes and usage data. |
| [AUDIENCE] | Name who will read or use the finished output. | The customer's operations lead and our account executive. |
| [AVAILABLE INPUTS] | List the documents, fields, facts, and constraints you can provide each time. | Account name, meeting notes, usage metrics, open issues, next meeting date. |
| [OUTPUT SPECIFICATION] | State the exact format, sections, length, and decision the output must support. | A 180-word email with wins, risks, owners, due dates, and next steps. |
| [QUALITY EXAMPLES] | Give a short example of an acceptable result and a result to avoid. | Good: names the owner and date for each action. Bad: calls usage 'strong' without a metric. |
How to use
- Paste the prompt and replace each bracketed field with the task you repeat.
- Keep source inputs in a consistent format, especially dates, metrics, and names.
- Check the proposed variables for overlap and remove any field you cannot supply reliably.
- Follow up with: "Create a blank fill-in form for these variables that my team can complete in two minutes."
Variations
Prompt audit
Use this when an existing reusable prompt gives inconsistent answers.
Audit this reusable AI prompt: [EXISTING PROMPT]. The recurring task is [TASK], and these are the inputs available each run: [INPUTS]. Identify ambiguous instructions, missing variables, conflicting rules, and places where the model may invent information. Return: (1) a table of issues and fixes, (2) a revised prompt with 2-6 uppercase square-bracket variables, and (3) three edge-case test inputs. Preserve instructions that work. Do not add requirements unsupported by [INPUTS]. Self-check that each revised variable is necessary and defined.
Variable schema
Use this when you need a structured input form for an automated workflow.
Design a variable schema for an AI workflow that does [WORKFLOW]. The output must be [OUTPUT]. Available source systems are [SYSTEMS]. Return a JSON-style field list with variable name, data type, required status, allowed values or format, example, and validation rule. Then provide a short prompt that consumes those variables. Use only fields needed for the output, distinguish user-entered from system-derived fields, and never treat missing data as a fact. Flag fields that need human approval.
Test cases
Use this when you want to test a prompt before your team adopts it.
Create a prompt test pack for [PROMPT OR TASK]. Its variables are [VARIABLES], and success means [SUCCESS CRITERIA]. Produce five compact test cases: ordinary, missing optional data, conflicting source data, unusually short input, and a high-risk edge case. For each, give filled inputs, expected behaviors, and failure signals. Do not invent domain facts; make generic placeholder content where needed. End with a pass/fail checklist that a reviewer can apply to each model response.
Tips
- Use variables for information that changes between runs, not for permanent instructions such as tone or output structure.
- Give each variable one authoritative source; two fields that can disagree, such as [CLIENT NAME] and a pasted brief, create silent errors.
- Specify the action for a blank optional field, such as omitting the section or labeling it "Not provided."
- Test with incomplete and contradictory inputs before sharing the prompt with a team.
FAQ
How many variables should a prompt have?
Usually two to six. More fields can be appropriate for a form, but a long variable list often hides unclear task instructions.
Can I use variables in ChatGPT, Claude, and Copilot?
Yes. Square brackets are a readable convention; replace them with real values before sending the prompt.
Should variables include example values?
Include one short example when a field has a strict format, such as dates, a table, or a list of approved options.