AI-Written Prompt Workflow
Specify the desired outcome and let the model draft its own prompt.
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 99%
Instead of improvising a short task prompt, first ask the AI to act as a prompt designer. Provide the task, the intended outcome, constraints, and—when possible—an example or description of the desired output. The model converts those inputs into a more complete prompt for itself or another model. Review that prompt for missing context, unsafe assumptions, and incorrect requirements, then execute it and refine based on the observed result. This separates prompt construction from task performance and uses the model’s familiarity with instruction patterns to overcome the user’s tendency to provide vague, underspecified requests.
Origin
Kristen Fraccia presented AI-written prompts as an underused technique, and Kieran Flanagan expanded it into prompt-specific onboarding assets.
Core principles
- 01Models can translate loose intent into stronger operating instructions.
- 02Desired outputs are easier to specify than perfect prompts.
- 03Prompt generation should precede task execution.
- 04A generated prompt still benefits from human review.
How to run it
- 1
Define the task
State what must be produced or decided in plain language.
Pro tip Name the intended user or audience.
Watch out A vague task will generate a polished but vague prompt.
- 2
Describe success
Specify the desired format, qualities, constraints, and example output if one exists.
Pro tip Include failure conditions as well as positive requirements.
Watch out Do not include sensitive data unnecessarily.
- 3
Request the prompt
Ask the model to turn the specification into a reusable prompt for the target model or task.
Pro tip Name the target model when its behavior matters.
Watch out Do not assume the generated prompt is automatically correct.
- 4
Audit the instructions
Check that the prompt preserves every requirement and does not invent unsupported context.
Pro tip Ask the model to list ambiguities separately.
Watch out Longer prompts are not inherently better prompts.
- 5
Execute and refine
Run the prompt, compare the output with the success criteria, and update weak instructions.
Pro tip Save prompts that work repeatedly.
Watch out Do not keep refining wording when the real issue is missing data.
In the wild
A marketer tells Claude that she needs a positioning statement, describes the audience and the exact output she wants, and asks Claude to create the full prompt before attempting the task.
→ The resulting task prompt is more detailed and useful than the marketer’s initial shorthand request.
Common mistakes
Skipping the output definition
The model cannot construct effective instructions when success remains undefined.
Trusting the generated prompt blindly
The metaprompt can still omit constraints or encode faulty assumptions that require human correction.
Is it for you?
Best for
It is best for users who know the result they need but struggle to formulate comprehensive instructions.
Not ideal for
It is not ideal when the user cannot define success or when hidden domain constraints require expert input.
From the transcript
“But I'll share another hack, which is just generalized, and I don't think people do this enough, which is using AI to write your prompts.”
“So if you start with a task and you're like, I want an output that looks like this, this is what I want to accomplish,…”
From the episode
The AI Workflow That Lets 50 People Do the Work of 500 ($2B Founder Reveals)