1. Name the buyer decision.
A prompt portfolio is the set of questions you use to observe AI answers over time. It is a research sample. It is not a list of every question your customers ask, and it is not a prompt-engineering portfolio for showing off generated work.
Start with one audience, one market and one decision. For example: an operations manager at a small UK agency choosing client-reporting software. The decision gives you a way to reject questions that produce interesting answers but have no bearing on the purchase.
Write the objective before the prompts: understand which products appear when this buyer asks about reporting across several clients. Do not start by filling a plan’s prompt allowance. A smaller set that someone reviews is more useful than a larger set nobody can explain.
2. Collect language and keep its source.
Look for questions in sales-call notes, support requests, internal site search and customer interviews. Use material you have permission to review. Remove names, email addresses and other private details before entering a question into a third-party AI service.
Search queries and keyword research can suggest a topic. They do not prove that people use the same wording in an AI conversation. Treat a team-written prompt as a hypothesis until you have evidence for its relevance.
For each candidate, record the original source, the buyer role, the decision stage and the reason to track it. If you cannot identify a decision the answer could inform, put the prompt aside.
3. Build a small, balanced first set.
This worked portfolio is illustrative. The agency, questions and reasons below are not observed customer data. It shows how six questions can cover different decisions without being six rewrites of the same keyword.
Keep discovery questions free of the brand name so you can observe which options appear without being prompted. Branded questions serve a different purpose: checking how an answer describes a product already under consideration. Review the two groups separately.
| Stage | Question to track | What to inspect |
|---|---|---|
| Discovery | Which reporting tools suit a UK agency with ten clients? | Which products enter the shortlist? |
| Requirement | Which agency reporting tools support separate client workspaces? | Is client separation explained? |
| Workflow | How can an agency combine reporting across several client accounts? | Which approaches and sources appear? |
| Comparison | How should an agency compare dedicated reporting software with spreadsheets? | Which tradeoffs shape the decision? |
| Branded review | What should I verify before using [brand] for client reporting? | Is the product described accurately? |
| Implementation | What should a ten-client agency check before changing reporting tools? | Are migration and access constraints covered? |
- Replace [brand] with the product being evaluated; do not submit the placeholder.
- Set the intended market and record the chosen platforms.
- Keep wording consistent between review periods.
- Give each question a topic and an owner for reviewing its answers.
4. Remove duplicates before the baseline.
“Best agency reporting software” and “top reporting tools for agencies” may test the same buying decision. Keep both only if the different wording is a deliberate part of your sample. Otherwise one topic gets extra weight simply because it has more variations.
Read a first answer for each question. Does it address the intended buyer and market? Does it return a product comparison when you wanted one? Replace an ambiguous question before you start comparing periods, and record why it changed.
A result with no AI answer is still useful evidence about that question on that platform. It is different from a failed collection. Check that distinction before deciding to retire the question.
5. Keep a baseline you can reproduce.
Save the exact prompts, platform selection, location, dates and any filters alongside the first review. Keep the answers and source URLs available so another person can check your conclusion.
For example, six prompts across three platforms create eighteen scheduled prompt-platform observations per day. That is an illustrative sample size, not eighteen people or eighteen independent buying decisions. Failed collection can reduce the number of recorded responses.
If a prompt is added or removed, annotate the change. A rising score after deleting difficult questions does not show that your brand became more visible for the original portfolio.
| Field | What to record |
|---|---|
| Objective and owner | The decision being studied and who reviews it |
| Prompt and source | Exact wording; interview, query or team hypothesis |
| Audience and market | Buyer role, location and relevant constraints |
| Sample | Platforms, dates and collection coverage |
| Observation | Answer text, detected mentions and cited URLs |
| Action | One proposed change, its owner and review date |
| Change log | Any edit to prompts, filters or platform selection |
6. Expand when the buyer need expands.
At the next review, separate changes in answers from changes in the portfolio. Read the answers behind the largest movements. Check whether a new product, source or interpretation explains the result before assigning work.
Add prompts when you enter a new market, discover a new requirement or need to investigate a specific gap. Retire questions that no longer represent the buyer, while preserving the reason and date in your working record.
Your first usable next step is to choose one buyer decision and collect six candidate questions with their sources. In Heading, use prompt research to build the set, then connect the findings to an objective and a task. This method is a suggested workflow, not a statistically representative survey or a guarantee of future visibility.
