Foundational language · chapter 1 of 17
Prompt and metaprompt
A prompt asks AI to perform a task. A metaprompt describes how good instructions should be built, evaluated or adapted before the task is performed.
1.1 · plain explanation
Two different levels of instruction
Prompt
A concrete instruction telling the model what to do now, which context it has and what the result should look like.
“Write an introduction to this guide for a beginner.”
Metaprompt
An instruction that creates, evaluates or improves other instructions. It defines the method before the concrete task is solved.
“First create a good writing prompt by clarifying audience, objective and format.”
Important: A metaprompt is not simply a longer prompt. The difference is what the instruction works on: the task or the way the task is instructed.
1.2 · the relationship
When the task is clear, a prompt is often enough
Use this when the task, context and quality requirements are already clear.
Use this when the instruction itself must be constructed, checked or adapted to several situations.
1.3 · side by side
The difference becomes clear when we compare the purpose
| Question | Prompt | Metaprompt |
|---|---|---|
| What does it work on? | The concrete task. | The instruction or method behind the task. |
| What does it produce? | Text, code, analysis or an action. | An improved prompt, template or decision rule. |
| When is it useful? | When the objective is already clear. | When several tasks need the same quality and structure. |
| Common mistake | Too little context or unclear requirements. | So much abstraction that it never leads to an actual task. |
1.4 · concrete example
The same website task at three levels
“Build a website.”
The model must guess the audience, message, content, format and what counts as complete.
The concrete page is described
Create an English landing page for a technical foundation course.
Audience: curious beginners.
Show two numbered chapters and one clear start button.
Use short paragraphs and accessible HTML.The method for creating the prompt is described
When someone requests a website:
1. Clarify the goal, audience and desired action.
2. Identify the required content and constraints.
3. Define the structure and acceptance criteria.
4. Then create a concrete implementation prompt.
5. Mark assumptions that still need testing.1.5 · demonstration
The lawnmower method: program a simple role with language
Here ChatGPT receives the role of a simulated height controller for a mower deck. The target height is set once. In each new message, the user supplies the measured height and the model answers with one permitted command.
Role, target, rules and response format
ROLE
You simulate the height controller of a lawnmower.
TARGET
target_mm = 40
tolerance_mm = 1
RULES
- Measured height below 39 mm: UP
- Measured height above 41 mm: DOWN
- Measured height from 39 to 41 mm: HOLD
- Missing, invalid or implausible value: STOP
RESPONSE FORMAT
Respond only with valid JSON:
{"command":"UP|DOWN|HOLD|STOP",
"target_mm":40,
"measured_mm":number_or_null}- You
34
ChatGPT
{"command":"UP","target_mm":40,"measured_mm":34} - You
43
ChatGPT
{"command":"DOWN","target_mm":40,"measured_mm":43} - You
40
ChatGPT
{"command":"HOLD","target_mm":40,"measured_mm":40}
The role, target and boundaries are set in the starter prompt.
Every user message contains the latest measured height.
A strict JSON format makes the response possible to validate before use.
The prompt gives the model a role. An agent also needs tools, working state, an action loop and controlled boundaries.
API and conversation state: Each model request is independent by default. A simple integration can send the instruction and the required history again. The Responses API can also chain responses or use a persistent conversation, so the application does not always have to rebuild the full history manually.
For a real machine: Never connect unrestricted model text directly to a motor. A local deterministic controller must validate the command, enforce minimum and maximum height, read limit switches, use a watchdog and choose a safe stop for invalid or missing responses. The model can demonstrate or suggest the high-level command; safety-critical control must remain local.
1.6 · try it yourself
Which level does the instruction work on?
“Summarise this report in five bullet points.”
Prompt. It directly requests a concrete result from a known document.
“Create a template that always clarifies the recipient and decision before summarising a report.”
Metaprompt. It shapes the instruction that will later be used for several reports.
“Check whether this prompt lacks context, format or quality requirements, and rewrite it.”
Metaprompt. The object it works on is another prompt.
1.7 · checklist
Before you send a prompt
- Is the result you want clear?
- Does the model have the context it actually needs?
- Are the format, boundaries and quality requirements described?
- Do you know what must be verified afterwards?
- Is this one task, or do you really need a reusable method?