Quick Answer
If you need AI output to sound local, do not start with a generic role-play prompt. Start with a small table of real regional constraints, turn that into 3 to 5 synthetic personas, then run the same copy or workflow through each persona before human QA.
As of 2026-04-21, the practical split is:
- use ChatGPT when you want one repeatable workspace for persona iterations, files, and review runs inside a single Project
- use Claude when the review depends on current Gmail, Calendar, or Drive context, plus private-project knowledge built from connected Drive files
- use Gemini when the source material and final asset already live in Google Docs or Sheets and you want to build and test personas close to the document itself
Synthetic personas are still a screening layer, not evidence of real user demand. Use them to catch localization mistakes early, then hand off the risky parts to real users or qualified local reviewers.
Regional Inputs Matter More Than Translation Alone
On 2026-04-21, NVIDIA published a Hugging Face tutorial for Nemotron-Personas-Korea. The release is a useful reminder that localized AI behavior improves when you ground the workflow in structured local data instead of treating translation alone as the problem.
The official tutorial currently contains one important number conflict. In the introduction, it says the dataset provides 6 million synthetic personas. In the dataset attribute table further down the same page, it says 7 million total personas from 1 million records x 7 personas each. The safer takeaway is not the exact top-line count. The stable part is the dataset shape and the method.
As checked on 2026-04-21, the same Hugging Face tutorial also lists these details:
| Item | What the official tutorial says |
|---|---|
| Persona fields | 26 fields |
| Geographic coverage | All 17 Korean provinces and 25 districts |
| Occupations | 2K+ categories |
| License | CC BY 4.0 |
| Grounding sources | Korean public institutions including KOSIS, the Supreme Court of Korea, and the National Health Insurance Service |
Source: Hugging Face tutorial, checked on 2026-04-21.
When This Workflow Is Worth Using
Use synthetic personas when the likely failure is obvious, repeated, and cheap to fix before launch.
Good fits:
- reviewing landing-page copy for a new country or region
- testing customer-support prompts before a local rollout
- checking whether onboarding assumes the wrong payment, document, or device default
- pressure-testing an agent for tone, workflow order, and trust signals
- comparing how one workflow lands with different occupations, age bands, or urban versus non-urban settings
Bad fits:
- replacing interviews or field research
- estimating demand or willingness to pay
- inferring protected traits you do not have a lawful basis to use
- generating fake testimonials or fake case studies
- regulated workflows where legal, medical, policy, or compliance accuracy needs qualified human review
Start With A Market Input Sheet
Do not begin with a paragraph prompt. Begin with structured inputs.
| Field | Example | Why it matters |
|---|---|---|
| Market | South Korea | Keeps legal and institutional context anchored |
| Region | Seoul, Busan, rural Jeolla | Local services and references change by region |
| Age band | 25-34, 55-64 | Device habits and language expectations shift |
| Occupation | office administrator, shop owner, student | Changes vocabulary and workflow priorities |
| Language register | formal, neutral, casual | Important for support and assistant tone |
| Device context | Android-first, desktop at work | Affects discoverability and task flow |
| Payment or process constraints | invoice, card, employer approval | Catches imported purchase assumptions |
| Goal | book appointment, compare tools, draft report | Keeps the persona tied to one real job |
If you cannot fill this table with public inputs, first-party research, or current internal documentation, stop. The persona will drift fast.
Pick The Right Workspace: ChatGPT vs Claude vs Gemini
The model matters less than where your context already lives.
| Tool | Best use in this workflow | Official capability checked on 2026-04-21 | Practical limitation to remember |
|---|---|---|---|
| ChatGPT | Repeated persona generation and review loops in one workspace | OpenAI says Projects keep chats, files, and project instructions together, and are available to all free and paid plans globally | If you need strict isolation, create the project with project-only memory instead of relying on default memory behavior |
| Claude | Persona reviews that need current Google Workspace context or project knowledge | Anthropic says Projects are available to all users, and Google Workspace connectors are available on Claude and Claude Desktop | Google Drive files can be added to private projects only; shared projects do not support that Drive-to-project flow |
| Gemini | Building and testing personas directly inside Docs or Sheets where the localized asset already lives | Google says Gemini is built into Docs, Sheets, Slides, Vids, and Forms, and prompt guidance is available for those apps | These features require an eligible Google Workspace or Google AI plan |
Sources checked on 2026-04-21: Projects in ChatGPT, What are projects?, Use Google Workspace connectors, Gemini in Docs, Sheets, Slides, Vids, & Forms, and Prompt guide for Gemini in Workspace.
Build The First Persona In ChatGPT
As of 2026-04-21, OpenAI says ChatGPT Projects can hold chats, files, and project instructions in one place. The same help page also says connectors are not currently supported inside Projects, so Project setup should assume direct uploads and project instructions rather than connected app sources.
Source checked on 2026-04-21: OpenAI Help Center, Projects in ChatGPT.
That makes ChatGPT the cleanest option when you want one reusable workspace for persona generation, audit, and review, and your core reference files can be uploaded directly.
Use this setup:
- create a new Project for one market or launch
- upload your market brief, style guide, policy notes, and any current copy draft
- open Project settings and add a short instruction block for this market
- if isolation matters, create the project with project-only memory
- if your source material lives in Drive, Slack, or another external system, export or paste the relevant material into uploadable files first
- run one persona-generation prompt per row in your input sheet
A compact Project instruction block:
You are helping with localization testing for one market.
Use only the files, source links, and structured inputs inside this project.
Do not invent personal history, protected traits, or unsupported preferences.
If a field is unknown, return "unknown".
Prioritize workflow constraints, tone expectations, and blocker risks.
Return structured outputs, not stories.
First persona prompt for ChatGPT:
Create one fictional synthetic persona for localization testing.
Use only these structured inputs:
- market: South Korea
- region: Seoul
- age_band: 25-34
- occupation: office administrator
- language_register: formal Korean in service contexts
- device_context: Android phone first, desktop during work hours
- payment_or_process_constraints: employer approval needed for some purchases
- goal: compare AI note-taking tools for team use
Rules:
- Do not invent personally identifiable information.
- If something is not supported by the input, mark it as unknown.
- Focus on workflow behavior, communication expectations, and local constraints.
- Return valid JSON only.
Return these keys:
- persona_label
- context_summary
- likely_tasks
- likely_constraints
- language_expectations
- local_workflow_risks
- review_questions
A cleaned persona output should look compact and auditable, not biographical. For example:
{
"persona_label": "Seoul office admin, formal service register",
"context_summary": "Works in an office environment, uses Android first, and may need employer approval before some software purchases.",
"likely_tasks": [
"compare AI note-taking tools for team use",
"review pricing and approval steps",
"check whether a tool works well on mobile before desktop follow-up"
],
"likely_constraints": [
"limited time during work hours",
"some purchases require internal approval",
"expects clear Korean service language"
],
"language_expectations": [
"formal tone in service interactions",
"avoid translated-sounding purchase instructions"
],
"local_workflow_risks": [
"consumer-style checkout assumptions",
"missing approval or invoicing details"
],
"review_questions": [
"Is the buying path clear for a team purchase rather than a personal subscription?",
"Does the copy explain approval or billing steps early enough?"
]
}
Why ChatGPT works well here: OpenAI's current help page says Projects are designed for repeatable workflows and let you add files plus project-specific instructions. That is enough to keep one market review loop stable across several rounds.
Build The First Persona In Claude
As of 2026-04-21, Anthropic says Claude Projects are available to all users and have their own chat histories, knowledge bases, and project instructions. Anthropic also says Google Workspace connectors are available on Claude and Claude Desktop, and that Google Drive files can be added directly to chats or to Files inside private projects, where they sync with the latest Drive version.
Sources checked on 2026-04-21: Claude Help Center, What are projects? and Use Google Workspace connectors.
Claude is the better fit when your synthetic persona should be grounded in current internal material rather than static uploads alone.
Use this setup:
- create a private Claude project for one market or workflow
- add product docs, support notes, or policy files to project knowledge
- if the source material lives in Google Drive, open the project, go to Files, click +, choose Drive, then add the relevant document
- run the persona prompt and ask Claude to label which parts are supported by the source material versus inferred
- keep the citation-rich review output in that same project
Prompt for Claude:
Create one fictional synthetic persona for localization testing.
Use only:
1. the structured inputs below
2. the project files and connected Google Drive documents available in this private project
Structured inputs:
- market: South Korea
- region: Seoul
- age_band: 25-34
- occupation: office administrator
- language_register: formal Korean in service contexts
- device_context: Android phone first, desktop during work hours
- payment_or_process_constraints: employer approval needed for some purchases
- goal: compare AI note-taking tools for team use
Requirements:
- No personally identifiable information.
- No unsupported biographical color.
- Separate supported facts from inferences.
- Cite any project file or connected document used.
- Return JSON only.
Return these keys:
- persona_label
- context_summary
- likely_tasks
- likely_constraints
- language_expectations
- local_workflow_risks
- review_questions
One reason Claude is useful in this workflow: the current connector docs say Claude responses include citations to the emails, calendar events, or documents used as sources. That makes assumption review easier when the persona is grounded in real internal material instead of a pasted brief.
Build The First Persona In Gemini
As of 2026-04-21, Google says Gemini is built into Docs, Sheets, Slides, Vids, and Forms. Google's current Workspace prompt guide also recommends four prompt parts for Gemini: Persona, Task, Context, and Format. That maps almost perfectly to synthetic persona work.
Sources checked on 2026-04-21: Google Docs Editors Help, Gemini in Docs, Sheets, Slides, Vids, & Forms and Learn about prompts for Gemini in Docs, Sheets, Slides, Vids & Forms.
Gemini is the most convenient option when the source brief and the asset under review already live in Google Workspace.
A simple setup looks like this:
- keep the market input table in Google Sheets or a briefing doc in Google Docs
- open Gemini inside that Sheet or Doc
- write the prompt using Google's four-part structure: persona, task, context, format
- if you need extra reference material, type
@and select the file you want Gemini to use, or work inside the document you already want to review - save each approved persona in its own tab, section, or linked doc
Prompt for Gemini in Docs or Sheets:
Persona: You are a localization testing assistant.
Task: Create one fictional synthetic persona for review of an AI software buying workflow.
Context:
- market: South Korea
- region: Seoul
- age band: 25-34
- occupation: office administrator
- language register: formal Korean in service contexts
- device context: Android phone first, desktop during work hours
- payment or process constraints: employer approval needed for some purchases
- goal: compare AI note-taking tools for team use
Format:
- Return valid JSON only.
- Use these keys: persona_label, context_summary, likely_tasks, likely_constraints, language_expectations, local_workflow_risks, review_questions.
- If a detail is not supported, write "unknown".
- Do not invent personal history or protected traits.
Why Gemini fits this use case: Google explicitly documents prompt structure for Workspace Gemini and shows that you can attach relevant files by typing @ and selecting them. If your localized draft already lives in Docs, you can generate the persona and test the copy without leaving the document workflow.
Run An Assumption Audit Before Any Review
This is the step most teams skip. It is also the step that keeps the persona from turning into fiction.
Use the same audit prompt in ChatGPT, Claude, or Gemini right after the first draft:
Audit the persona you just created.
For each field, label it as one of:
- directly supported by the input
- reasonable inference
- unsupported guess
Then rewrite the persona so that unsupported guesses are removed.
Return:
1. an assumption audit table
2. a cleaned persona JSON
If you are using Claude, also ask for citations to any project file or connected source used in the audit. If you are using Gemini inside Docs, keep the audit directly below the persona so the edit trail stays visible.
Test One Real Asset Across 3 Persona Types
Once the cleaned persona exists, turn it into a test harness.
For a landing page or pricing page:
Act as the synthetic persona below.
Review this landing-page draft for clarity, trust, and workflow fit.
Evaluate only:
- what is immediately clear
- what feels imported from another market
- what sounds unnatural in local service language
- what step would likely block conversion
- what proof or reassurance is missing
Return:
- first impression
- confusing phrases
- risky assumptions
- likely objections
- one rewrite suggestion
- one blocker question that still needs a human reviewer
For a support or onboarding flow:
Act as the synthetic persona below.
You are trying to complete this task:
"Reschedule a clinic appointment through an AI assistant."
Review whether the flow matches local expectations.
Flag:
- wrong level of politeness
- wrong institutional assumptions
- missing local documents or steps
- translated-sounding terms
- moments where a human handoff is needed
- one blocker question that should be checked by a local reviewer
Do not stop at one persona. A first pass usually needs these three:
| Persona type | What it catches best | Typical failure signal |
|---|---|---|
| Mobile-first urban worker | speed, app flow, missing next steps | "I still do not know what to click first." |
| Older service user | jargon, trust language, formal tone mismatches | "This sounds translated, not native." |
| Small-business operator | approval chains, invoicing, admin friction | "This workflow assumes a consumer purchase." |
Keep The Library Small And Dated
Most teams do not need a giant persona catalog. They need a small one they can trust.
A good default is 3 personas for a first pass and 5 to 8 personas for a maintained market library.
Store each persona with:
- the input table that produced it
- the assumption audit
- the workflow it is meant to test
- the date local facts were last checked
- the tool and workspace used to create it
That last field matters. If a persona was generated from Claude with current Drive files, it is not equivalent to one generated from a static ChatGPT project three weeks earlier.
Common Mistakes
Treating synthetic personas like real respondents
Do not write things like "Korean customers told us" if the finding came from a synthetic review pass.
Write the claim as a workflow finding instead:
Our synthetic review pass on 2026-04-21 flagged this as a likely localization risk. Validate it with real users or a qualified local reviewer before rollout.
Letting the model invent identity details
Favorite foods, family history, and personality color usually make the test worse, not better.
Cut the persona back to task, constraints, register, region, and environment.
Testing tone but not blockers
A localized flow can sound fluent and still fail on the wrong ID type, approval order, business-hours assumption, or escalation path.
Every review run should end with at least one blocker question.
Keeping stale personas for too long
Local workflows change. Payment defaults change. Product docs change. Plan limits change.
If the source material moved, rerun the persona.
FAQ
Are synthetic personas enough for market entry decisions?
No. They are useful for finding obvious localization mistakes, not for proving demand or replacing interviews.
Which tool should most teams start with?
Start with the tool that already sits closest to your working materials.
- pick ChatGPT if you want one reusable review workspace with files and project instructions
- pick Claude if current Drive, Gmail, or Calendar context changes the answer materially
- pick Gemini if the draft and source material already live in Docs or Sheets
Can I use this in regulated industries?
Only as a draft-review layer. Legal, medical, policy, and compliance claims still need current local verification by qualified humans.
What is the biggest prompt mistake?
Asking the model to make the persona feel vivid instead of making the persona easy to audit.
Verification Note
Verified on 2026-04-21 against official sources.
Checked items:
- OpenAI Projects behavior, availability, file uploads, project instructions, connector limitations inside projects, and memory options
- Anthropic Projects availability, Google Workspace connector behavior, and the private-project requirement for adding Drive files to project knowledge
- Google Workspace Gemini availability and Google's current prompt guidance for Docs and Sheets
- NVIDIA's Hugging Face tutorial for Nemotron-Personas-Korea, including the currently inconsistent 6 million versus 7 million top-line count and the attribute-table details for fields, coverage, occupations, and license
Official sources:
- OpenAI Help Center: Projects in ChatGPT
- Claude Help Center: What are projects?
- Claude Help Center: Use Google Workspace connectors
- Google Docs Editors Help: Gemini in Docs, Sheets, Slides, Vids, & Forms
- Google Docs Editors Help: Learn about prompts for Gemini in Docs, Sheets, Slides, Vids & Forms
- Hugging Face: How to Ground a Korean AI Agent in Real Demographics with Synthetic Personas