How to Write a Synthetic Research Brief That Can Be Checked
A synthetic research brief should define the decision, relevant audience, allowed evidence, exact stimulus and the questions a simulation will explore. It should also specify what the output cannot establish and how it will be checked with people or records. A persuasive description of the customer you hope to have is not a substitute for that brief.
Our guide to synthetic personas explains the representations involved. This article supplies a practical way to set their task without silently turning assumptions into customer facts.
Begin with a decision that could change
Write a sentence naming the decision and owner: “The product team will decide whether to revise the collection offer before a human pilot.” Then name the uncertainty: “We do not know whether the eligibility conditions are understandable.” This is more useful than “generate insights into our ideal customers”.
Separate the business decision from the simulation's immediate output. The business may eventually decide whether to launch. The simulation may only help prepare questions for a pretest. Stating both prevents a preparatory exercise from being presented later as launch validation.
Golder and colleagues (2023) describe an empirics-first approach grounded in real marketing phenomena and data. This is a methodological argument, not an endorsement of synthetic evidence. Its relevance to the brief is to start from an observable problem and specify what information is missing.
Define the audience in decision-relevant terms
Identify the category, market, access conditions and relevant behaviour. Distinguish criteria supported by evidence from attributes added for exploration. Include exclusions and explain them. Do not narrow the audience to people already enthusiastic about the proposed solution.
For a growth question, ask whether you need perspectives beyond frequent buyers of your brand. Trinh, Dawes and Sharp (2024) distinguish current buying contributions from growth headroom and find substantial opportunity among light and non-brand buyers in their examined data. This does not make every non-buyer equally promising; it makes a loyal-customer-only brief something to justify.
Keep category purchasing separate from brand purchasing. A light buyer of your brand can be a heavy buyer of the category. A brief that calls both “low engagement” loses a distinction relevant to the decision before the simulation begins.
Use a brief with explicit evidence labels
The following template is original and can be adapted. It is a research planning aid, not a guarantee of valid synthetic responses.
| Field | What to write |
|---|---|
| Decision | The action under consideration, the owner and what could change it |
| Immediate task | The bounded question this simulation will explore |
| Audience | Relevant category, market, behaviour and access conditions |
| Known information | Source-backed facts, with source and date |
| Assumptions | Conditions being explored that have not been established |
| Stimulus | The exact offer, image or text to evaluate, including relevant constraints |
| Questions | The wording, order and response options to use |
| Allowed inputs | What reference material the process may receive |
| Output | The requested format and how evidence status should be labelled |
| Exclusions | Claims the exercise must not present as measured facts |
| Evaluation | The human or behavioural reference and criteria for judging usefulness |
| Handoff | The next research action, owner and record to retain |
Keep the brief short enough to inspect but complete enough to reproduce the task. A long biography can be less useful than a clear statement that the offer is unavailable outside a certain area.
Worked example: a collection service
Imagine a shop considering weekly collection of pre-ordered household essentials. The immediate task is to identify possible misunderstandings in the offer description before a human pretest. The eventual business decision is whether to run a paid pilot.
Known facts include the proposed collection hours, fee and order deadline. The audience is households able to use the collection point who buy relevant products. The team has not established how many want the service or what they would pay; those remain open questions.
The stimulus states what is included and excluded. Questions ask the simulation to paraphrase the offer, identify missing information and propose reasons the service might be impractical. The requested output distinguishes interpretation of supplied facts from additional hypotheses.
The exclusions say not to present generated stories as actual shopping experiences, generated frequencies as market estimates or simulated purchase intention as validated demand. The handoff is a revised description and a question list for people eligible for the real pilot.
This is a fictional brief. The example shows how each field contributes to an inspectable task; it is not evidence that the service would succeed.
Remove assumptions that contain the desired answer
Compare “time-poor households who value the convenience of our collection service” with “households responsible for purchasing these products, with different schedules and existing shopping arrangements”. The first embeds both a problem and a favourable interpretation of the solution.
Similarly, “explain why this price is reasonable” requests a justification. “What information would be needed to assess this fee?” leaves room for uncertainty. If the goal is to challenge a proposition, say so explicitly rather than treating the resulting argument as an unbiased population response.
Do not instruct a persona to have a precise purchase history and then report that history as a finding. Scenario facts can be useful inputs, but they are part of the setup. Their appearance in the output is not independent confirmation.
The persona-generator guide provides a practical distinction between observed characteristics, interpretations and fictional detail.
Ask for an output that preserves uncertainty
Request separate fields for the answer, supporting supplied information, assumptions introduced and questions needing external checks. Do not require a confident recommendation when the input does not support one. An “insufficient information” response can be more useful than a polished invented explanation.
For concept work, preserve the exact wording of possible misunderstandings so a researcher can inspect whether the stimulus invites them. For creative work, distinguish describing an element from predicting its effect on people. For questionnaire work, identify the item that creates an ambiguity and propose a revision for review.
Sarstedt and colleagues (2024) discuss the value of silicon sampling in upstream pretesting and pilots. The brief should make that role operational: the output improves the next research step rather than pretending the step has already been completed.
Keep enough information to reproduce the task
Save the final brief, exact stimulus, questions, date and available model information alongside the outputs. Record revisions and why they were made. If settings are unavailable, state that limitation rather than filling in assumed values.
Ong (2024) highlights the importance of reporting prompts, procedures and relevant model settings. In applied work, these records also help distinguish a changed brief from a changed generating process when results differ.
If you are validating against a completed human study, separate its evaluation answers from the inputs. Do not let an expected conclusion enter the brief and then count its repetition as predictive success. Follow the validation protocol for held-out comparisons and baseline checks.
Close the brief with a research handoff
Before running the simulation, decide who will review its suggestions and what happens next. A useful handoff might contain a revised concept description, an unresolved-assumptions list and a human-pretest questionnaire. Assign an evidence requirement to each consequential claim.
Track what the simulation missed as well as what it anticipated. If human research raises an issue absent from the generated output, retain it as a failure of coverage. If a predicted problem never appears, do not continue presenting it as an established customer barrier.
Prepare the brief before generating the personas. Adapt the fields above, review the concept-testing questions, and use SynthFolk's qualitative workflow for a labelled rehearsal. Check current pricing and reserve the human or behavioural research needed for the final decision.
Sources and editorial method
The template and collection-service example are original applications. The research cited below supports specific distinctions about problem definition, audiences and reproducibility. It does not validate the template or demonstrate SynthFolk's predictive performance.
- Golder, P. N., et al. (2023). Learning from Data: An Empirics-First Approach to Relevant Knowledge Generation. Journal of Marketing, 87(3), 319–336. DOI: 10.1177/00222429221129200.
- Trinh, G. T., Dawes, J., & Sharp, B. (2024). Where is the brand growth potential? An examination of buyer groups. Marketing Letters, 35, 95–106. DOI: 10.1007/s11002-023-09682-7.
- Sarstedt, M., Adler, S. J., Rau, L., & Schmitt, B. (2024). Using large language models to generate silicon samples in consumer and marketing research: Challenges, opportunities, and guidelines. Psychology & Marketing. DOI: 10.1002/mar.21982.
- Ong, D. C. (2024). GPT-ology, Computational Models, Silicon Sampling: How should we think about LLMs in Cognitive Science? arXiv:2406.09464, version 1.