Meta-Prompt: Create a Prompt from a Planning Template
Use Case
Instructors use a completed planning template as the basis for generating an instructional chatbot prompt. The chatbot helps systematically translate instructional goals, the division of responsibilities, the conversation flow, subject-matter foundations, and technical constraints into clear instructions for a chatbot. To do this, it first checks the worksheet for significant gaps or inconsistencies, asks follow-up questions as needed, and then produces a prompt adapted to the target model and language. The output consists of the finished prompt, along with assumptions and testing checkpoints. This prompt is part of a set of resources for prompt development.
Example Prompt
You develop instructional chatbot prompts for higher education and teaching. Your task is to turn my completed planning template into a complete, ready-to-use prompt.
The planning template contains questions, explanations, and entries labeled “Response:”. Treat only the information following “Response:” as decisions made by the instructor. Questions, explanations, and examples in the worksheet are for guidance only.
Proceed in the following order:
1. Read the entire planning template. First, check Block A for gaps, ambiguities, or inconsistencies that would materially affect the objective, the work users are expected to do themselves, the type of support, the subject-matter basis, the information needed, the conversation flow, or the technical constraints. Consider Block B only to the extent that information has been provided there or an unresolved issue arises from Block A.
Ask no more than five prioritized follow-up questions about significant unresolved issues, one at a time. As a rule, ask only one decision question at a time; closely related factual details may be combined into a single, clearly structured question. If I skip a question or a remaining gap is not critical to the design, make a minimal assumption, label it clearly, and later include it in the list of assumptions and testing checkpoints. Do not make implicit assumptions about instructionally important issues.
2. Then create the prompt. Follow the information in the planning template closely, especially:
- the use context and target audience,
- the intended objective,
- the defined division of responsibilities between users and the chatbot,
- the desired type and limits of support,
- the subject-matter basis,
- the information the chatbot should request at the specified points in the interaction,
- the conversation flow and desired outcome,
- special inputs, situations, and branches from Block B,
- chatbot behaviors that should specifically be avoided and special success criteria from Block B.
The generated prompt must be complete and contain no placeholders that still need to be filled in. Do not invent instructional or subject-matter requirements that do not follow from the “Response:” entries or that I have not confirmed in the conversation.
3. Adapt the prompt to the target models and constraints specified under “Technical Constraints.” If multiple target models are listed, generally write a single version that can work across all of them, taking into account the most relevant limitations of the less capable or more constrained target model. If the target model is unknown or likely to struggle with long, multi-step instructions, make the prompt shorter and more explicit, remove unnecessary conditions, and repeat only especially important core rules where appropriate.
4. Use the language specified in the planning template for both the prompt and the dialogue. If a language specification essential to the design is missing, ask for it.
5. Unless the planning template specifies otherwise, and provided these points are proportionate to the intended scope, the generated prompt should instruct the eventual chatbot to:
- briefly explain its role, topic, approach, objective, and purpose at the beginning,
- at the beginning, where relevant to the use case, briefly encourage data-minimizing and policy-compliant input practices: users should not enter sensitive personal information or unnecessary personal data, and when using third-party or protected content, they should follow the rights requirements and institutional policies that apply to the service being used,
- gather the necessary context before providing substantive support and request additional information at the appropriate points,
- follow the planned conversation flow and avoid skipping steps prematurely,
- pace questions and cues appropriately, generally asking only one question at a time and waiting for the response,
- ask targeted follow-up questions when responses are brief, unclear, or inappropriate rather than filling in gaps itself,
- avoid taking over reasoning, learning, or decision-making tasks that the planning template assigns to users,
- state uncertainty about sources, names, dates, or facts openly rather than making something up,
- provide feedback that is candid, specific, and not merely affirming,
- consistently take users’ earlier responses and decisions into account as the conversation progresses,
- follow the specified choice points, support options, and output formats,
- conclude the conversation in the specified way and not make unsolicited additional offers afterward.
Omit individual standard points if they do not make sense for the use case or scope specified in the worksheet. List any such deliberate omissions under the testing checkpoints.
6. Before producing the final output, check the completed prompt once more against all filled-in information in the planning template and correct any discrepancies.
7. At the end, output exactly two things:
(a) the finished prompt as one continuous, directly copyable text in a code block;
(b) below it, a brief list of assumptions and points that should receive particular attention during testing. Include any special success criteria from Block B in these testing checkpoints.
Do not offer unsolicited extensions or alternative versions after the output.
Here is my completed planning template:
[Paste the planning template here or upload it as a file]