Custom persona prompts you can actually use
A persona prompt is useful when it gives ChatGPT a clear job, audience, output shape, and boundary. Copy one of these, replace the bracketed context, and refine it with one real example from your work. They are starting points, not hidden expertise.
1. Customer-support reply editor
Act as a support-reply editor for [company]. Rewrite my draft in a calm, direct tone. Preserve only facts in the draft. Do not invent policy, promise a refund, or claim an action was completed. If information is missing, list the one question a human should answer before sending.
2. Help-center article planner
Act as a help-center editor. Turn this repeated customer question into an article outline: title, who this applies to, ordered steps, exceptions, and when to contact support. Flag claims that need a source rather than making them up.
3. Product-release explainer
Act as a product marketer writing for existing customers. Explain this release in plain language: what changed, who benefits, how to start, and what did not change. Keep the answer under 250 words and ask for missing facts.
4. Meeting-notes analyst
Act as an operations analyst. From these notes, produce decisions, owners, deadlines, open questions, and risks. Do not turn suggestions into decisions. Quote the line that supports each decision.
5. Research skeptic
Act as a skeptical researcher. Separate this proposal into claims, evidence supplied, assumptions, and questions that would change the decision. Do not add sources I did not provide.
6. SQL review partner
Act as a careful SQL reviewer. Explain what this query changes or returns, identify destructive or expensive parts, suggest a read-only check first, and provide a rollback question. Do not execute or claim to have executed SQL.
7. Interview question designer
Act as an interview designer for [role]. Create five behavioral questions, the skill each tests, and a follow-up question that asks for evidence. Avoid questions that assume a candidate's background.
8. Plain-language policy editor
Act as a plain-language editor. Rewrite this policy for a customer who needs to act on it today. Keep the original meaning, label exceptions clearly, and flag ambiguous or conflicting terms instead of resolving them yourself.
What a persona can and cannot do
A persona tells an AI how to sound and what role to play. “Be concise, patient, and direct” is useful. So is “explain published return policy, but do not approve exceptions.” Neither instruction gives the AI the current policy, verifies a customer's identity, or routes an unresolved case. Use Custom Instructions when you want persistent preferences in ChatGPT, and keep company policy in a source your team maintains.
Keep durable facts in the source your team maintains, such as a help center or policy document. Keep the persona focused on behavior: tone, job, allowed sources, prohibited claims, and escalation conditions. This makes updates easier to review and avoids hiding a policy inside a long prompt.
How to use a 300-persona list without collecting prompts
The list is useful when it helps you name the work you want done. Try a few roles against the same small task, then keep the one whose output needs the least editing. These are starting prompts, not facts or permission to act.
| Persona pattern | Good first use | Starting prompt |
|---|---|---|
| Researcher | Turn sources into open questions | “List claims, evidence, and what needs checking.” |
| Editor | Improve a draft without changing meaning | “Mark unclear sentences and propose plainer replacements.” |
| Tutor | Explain an unfamiliar concept | “Teach this in steps; ask one check question.” |
| Analyst | Compare stated options | “Make a decision table and name missing information.” |
| Support triager | Sort a fictional request | “Classify the request, name the source needed, and escalate if facts are missing.” |
Use a one-off prompt for an experiment. Use Custom Instructions for stable preferences across chats; they can be edited or removed. Use a custom GPT for a reusable ChatGPT experience with configured instructions, knowledge, and selected capabilities. None replaces identity checks, current policy, or human accountability in a customer workflow.
Build one persona around a real job
Start with a narrow support job, such as answering delivery-window questions. Name the source owner. State what the teammate may say, what it must ask for, and when it sends the conversation to a person. Test a normal question, a request with missing information, an exception, and a human request. Inspect the answer, source, ticket state, and handoff, not just whether the prose sounds friendly.
The same applies to a brand voice. Choose a tone that a customer can understand and a human agent can maintain. Do not use a clever character voice if it makes a refund, safety, or account question harder to understand.
Review an eesel teammate from the terminal
eesel's CLI works with the same teammate and workspace as the dashboard. A support lead can ask Claude Code, Codex, or Cursor to check whether its delivery persona asks for missing facts without sounding dismissive. The coding agent reads the teammate's instructions and JSON responses, then proposes a precise wording change. A person or script can run the same review.
It requires Node.js 18.17 or newer. With the workspace owner's permission, inspect the current instructions, connected sources, and action permissions before testing. Use a teammate whose permissions are appropriate for the review. Chat runs real, billed workspace work: a fictional prompt is not a sandbox, and asking for no actions is not an access control. The approvals command lists held actions; it does not put every command or chat behind approval. Replace brand-support with your intended teammate.
npx @eesel/cli instructions --agent brand-support
npx @eesel/cli approvals --agent brand-support
npx @eesel/cli new --name "delivery-persona-review" --agent brand-support
npx @eesel/cli chat "Using this fictional request—an international order has no tracking number—draft a short triage response, name the missing fact, and propose human handoff. Do not send a message or change a ticket." --agent brand-support
npx @eesel/cli activity --agent brand-support
Check the JSON response and activity for the expected result: it should ask for the missing tracking information rather than invent it, identify the delivery-policy source it needs, and propose a human handoff. Have the coding agent show the current instruction beside its suggested replacement, such as asking one clear question before explaining the handoff. The owner reviews that exact change before it is applied, using command help and --dry-run where supported to inspect the proposed write.
Then start fresh conversations for the missing-tracking case, a normal delivery question, and an explicit request for a person. Compare factual accuracy and handoff behavior as well as tone. Verify delivery and routing in the actual helpdesk channel separately; a good CLI reply alone does not prove that customers reach the right queue.
When a prompt library is not enough
Prompt libraries are useful for drafting, research, editing, and structured thinking. For a customer-facing support job, eesel is an AI helpdesk teammate with approved knowledge, a helpdesk workflow, and reviewable activity. Its dashboard and CLI operate the same workspace, so the team can inspect knowledge, instructions, activity, and held actions before expanding the role. Learn more or try eesel.
Frequently Asked Questions
What are custom personas for ChatGPT?
They are prompts that set a role, tone, or style for a ChatGPT conversation. They can be useful for exploration, but they do not automatically provide company knowledge or workflow access.
Do I need 300 custom personas for ChatGPT?
No. A smaller prompt library is more useful when every prompt names a real job, an output format, and a boundary. Start with one or two that match your work, then refine them against real examples.
How should a support persona be written?
State the job, tone, allowed sources, actions it may take, and when it must hand a case to a person. Keep policy facts in maintained knowledge, not a giant prompt.
Can a persona safely answer every question?
No. It should decline or escalate requests outside its approved sources, identity checks, or action permissions.
How can a team test an AI persona?
Use representative support cases, including missing information and human requests, and review the exact response, sources, action record, and handoff result.
