Goal-Example-Data Prompt Formula
Give the model one job, show success, and supply the raw inputs
- Difficulty
- Starter
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 96%
Define the model's job as one clear deliverable, provide an example that demonstrates what good looks like, and then supply the raw information needed to produce a new version. This combination gives the model an objective, an output pattern, and task-specific evidence. Start by copying the information manually so the prompt can be tested at negligible cost. Compare results across several inputs and revise only where inconsistency appears. Rather than treating prompt wording as specialist engineering, ask a capable frontier model to critique and rewrite the prompt itself. This meta-prompting loop accelerates improvement while keeping the practitioner responsible for judging quality. Once the prompt reliably produces useful work, it can become the core capability inside a larger retrieval or automation system.
Origin
Ethan Dewal began meta-prompting while building Asana's call-preparation prompt. After seeing AI generate strong personalized emails, he realized the same model could help write the instructions for other AI use cases.
Core principles
- 01Assign one clear job to the model
- 02Demonstrate the desired output
- 03Supply relevant raw information
- 04Test usefulness before optimizing language
- 05Let a capable model improve the prompt
How to run it
- 1
Choose one deliverable
Describe the model's job as a single concrete output, such as a call-preparation document. Avoid mixing unrelated objectives into the same instruction.
Pro tip Phrase the goal so a reviewer can tell immediately whether the model completed it.
Watch out Multiple competing goals make failures difficult to diagnose.
- 2
Show what good looks like
Provide a representative example of the desired structure, content, depth, and tone. Use an example that an experienced practitioner considers genuinely strong.
Pro tip Remove sensitive information while preserving the example's useful structure.
Watch out A weak example teaches the model to reproduce weak work.
- 3
Supply raw information
Insert the account, customer, product, or meeting information needed for the task. During early testing, manual copy-and-paste keeps the experiment cheap and flexible.
Pro tip Separate instructions, example content, and raw inputs with clear labels.
Watch out Do not expect the model to know private or current information that was never supplied.
- 4
Evaluate across cases
Generate outputs for several representative inputs and compare them with the standard demonstrated by the example. Identify recurring omissions, hallucinations, or formatting problems.
Pro tip Include easy, typical, and difficult cases in the test set.
Watch out Do not judge consistency from a single impressive response.
- 5
Use meta-prompting
Ask a capable model to improve the prompt based on the desired goal and observed failures. Review its changes, then test the revised prompt against the same cases.
Pro tip Include examples of failed outputs so the model can target specific weaknesses.
Watch out Do not accept a longer rewritten prompt merely because it sounds more technical.
- 6
Operationalize only after reliability
When outputs are consistently useful, save the prompt as a specialized assistant or embed it in a workflow. Continue monitoring quality as models and inputs change.
Pro tip Retain a small regression set of representative inputs and expected characteristics.
Watch out Model updates can alter behavior even when the prompt remains unchanged.
In the wild
A representative tells the model that its job is to create a call-preparation document, supplies an example of a strong document, and pastes raw account information. After checking several results, the representative asks Claude to improve the prompt and retests the revision before sharing it with colleagues.
→ A useful specialized assistant is produced without conventional software engineering.
A seller defines the job as writing one personalized outbound email, provides a successful message as the pattern, and supplies the prospect's role, company, and relevant product information. The result is checked for factual accuracy and fit before the prompt is reused.
→ High-quality one-to-one messaging can be produced much faster than manual research and drafting.
Common mistakes
Overengineering the prompt
Modern frontier models often need clear instructions more than elaborate tricks. Complexity should address observed failures rather than imagined ones.
Omitting the example
A broad instruction leaves quality and structure underspecified. A strong example makes the target concrete.
Trusting the first output
One good result does not prove consistency. Test multiple representative inputs before turning the prompt into a shared workflow.
Is it for you?
Best for
It is best for repeatable document-generation tasks where users can identify a good example and gather the relevant inputs.
Not ideal for
It is not ideal for tasks requiring deterministic calculations, authoritative live data, or actions that need strict external controls.
From the transcript
“your job is to generate me a call preparation document here is an example of what a a great call preparation document looks like for…”
“it really is that simple with AI”
“if it can write me a really great hyper-personalized email why could it not write me a prompt”
From the episode
He Automated His Sales Job With Ai… So His Boss Promoted Him