Prompt Design for Human-Centred AI
本页目录· 8 条
Core argument
A prompt that merely completes a task is the article’s “wire” prompt. A well-designed “cloth” prompt understands the user’s context and supports a usable experience. Model capability does not remove the need for product and interaction design.
Because model outputs are probabilistic, prompts cannot be treated like deterministic code. They require iterative testing of both user behaviour and model behaviour.
Six prompt-design layers
- Flow — the sequence that helps the user reach an outcome.
- Role — the stance/capability the model should adopt.
- Mission — the concrete job and success condition.
- Guidelines — behavioural constraints and important rules.
- Examples — representative inputs and desired responses; include edge cases, not only the happy path.
- Output parameters — structure, format, length and required fields.
Two-sided test
Observe the user
- Do they know what to ask?
- What context do they naturally provide or omit?
- When are they confused, over-trusting or dissatisfied?
- Can they correct the system?
Observe the model
- Does it preserve important user wording and detail?
- Does it behave consistently across representative and adversarial cases?
- Does it make unsupported claims or unsafe recommendations?
- Does the output meet the required structure?
Use findings to revise the flow, examples and guidelines—not just add more words.
FIT5120 prompt-spec template
- User/job:
- Context available:
- Desired outcome:
- Conversation flow:
- Role:
- Mission:
- Guidelines / prohibited behaviour:
- Positive examples:
- Edge/adversarial examples:
- Output schema:
- Human review / escalation:
- Success and failure measures:
- Version and test evidence:
Design cautions
- Do not place secrets or unnecessary personal data in prompts.
- Treat prompt text as versioned product logic with review and testing.
- Include transparent limits and a safe fallback.
- Distinguish an appealing interaction from a reliable, ethical product.