Create a production-ready prompt specification for an AI agent that handles [AGENT JOB] for [END USER OR TEAM]. The agent will receive [INPUTS], may use [ALLOWED TOOLS OR SOURCES], and must produce [REQUIRED OUTPUT]. Its operating constraints are [POLICIES AND BOUNDARIES]. Write the specification in a copyable format with these sections: 1. Role and success condition: define the agent’s narrow responsibility, the user it serves, and what a successful result can be observed to contain. 2. Input contract: list required versus optional inputs, accepted formats, validation rules, and what the agent should do when an input is missing, stale, contradictory, or outside scope. 3. Operating procedure: give a numbered workflow from intake through evidence gathering, reasoning, draft creation, validation, and handoff. State which sources take precedence when they conflict.
4. Tool rules: for each allowed tool or source, define when it may be used, what it may not change, and how results should be cited or summarized. Do not authorize actions beyond the stated boundaries.
5. Output contract: specify exact headings, tables, fields, length limits, and what uncertainty looks like in the final answer.
6. Guardrails and escalation: define prohibited actions, facts the agent must never invent, privacy handling, approval gates, and clear triggers for asking the user or escalating to a human.
7. Test cases: provide three compact example inputs, including one missing-data case and one edge case, with expected agent behavior rather than fabricated results.
Make instructions executable, not aspirational. Separate known facts from inferences, and require the agent to cite or identify the input that supports material conclusions. Do not claim the agent can browse, access systems, or perform actions not included in [ALLOWED TOOLS OR SOURCES].
Before answering, check for contradictory instructions and remove them; then confirm every external action has an explicit approval rule and every required output field has a source or a missing-data behavior. Ask up to 3 clarifying questions only if a required input is missing.Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [AGENT JOB] | Describe the repeated task the agent should perform in a narrow, concrete way. | triage incoming vendor security questionnaires and prepare a first-pass response matrix |
| [END USER OR TEAM] | Name the people who will use the agent and the decision or workflow it supports. | the security operations manager and sales engineering team |
| [INPUTS] | List the documents, forms, messages, data fields, and context the agent will receive. | a customer questionnaire, the approved security knowledge base, the product architecture summary, and prior approved answers |
| [ALLOWED TOOLS OR SOURCES] | List systems, files, APIs, or knowledge sources the agent is actually allowed to use. | read-only access to the approved knowledge base and a shared response-matrix template; no email sending or document publishing |
| [REQUIRED OUTPUT] | Describe the result the agent must return or prepare, including who uses it next. | a response matrix with answer draft, source reference, confidence level, owner for review, and unresolved question |
| [POLICIES AND BOUNDARIES] | List privacy, approval, compliance, tone, scope, and action limits the agent must follow. | do not make security commitments, disclose confidential architecture details, or mark an answer approved; escalate legal, privacy, and contractual questions |
How to use
- Define one repeated job and its handoff; avoid combining research, approval, execution, and support into one vague agent.
- List tools the agent can truly use, including whether it has read-only or write access.
- Run the test cases against real but non-sensitive examples before letting the agent touch a live workflow.
- Follow up with: “Convert this specification into a system prompt under 1,500 words, preserving every approval gate and output field.”
Variations
Research agent
Use for an agent that gathers evidence but does not make external changes.
Write an AI research-agent prompt for [RESEARCH QUESTION] serving [AUDIENCE]. It can use only [ALLOWED SOURCES] and must return [DELIVERABLE]. Define a search plan, source-quality hierarchy, note-taking schema, citation requirements, how to distinguish quotation, fact, and inference, and a stop condition. Require it to report gaps and conflicting evidence rather than resolve them silently. It must not contact people, purchase access, alter records, or present unverified claims as findings. Include checks for source recency and conclusion-to-source support.
Support triage
Use for an agent that classifies and drafts customer-support responses.
Create a support-triage agent prompt for [PRODUCT] handling [CHANNELS] for [CUSTOMER TYPE]. Inputs include [TICKET DATA], and the output is [RESPONSE OR ROUTING FORMAT]. Define severity levels, authentication and account-data limits, when to ask a clarifying question, routing rules, approved knowledge sources, and when the agent must avoid troubleshooting or promise no timeline. The agent may take these actions only: [ALLOWED ACTIONS]. Require it to label confidence and cite the relevant policy or article. Include escalation triggers for security, billing, legal, safety, and outage reports.
Content reviewer
Use for an agent that checks drafts against a defined editorial standard.
Create an editorial-review agent prompt for [CONTENT TYPE] intended for [AUDIENCE]. It receives [DRAFT AND SOURCES] and returns [REVIEW FORMAT]. Define review criteria for factual support, audience fit, structure, claims, tone, accessibility, and required disclosures. Require comments to quote or locate the issue, explain why it matters, and suggest a revision without rewriting unsupported facts. State which changes require human approval: [APPROVAL RULES]. The agent must preserve authorial choices that do not violate the criteria and flag missing sources or high-risk claims.
Tips
- Give an agent a narrow decision boundary; a reliable intake-and-draft agent is easier to test than an agent told to “handle customer operations.”
- Specify what happens with absent or conflicting inputs, because that behavior determines whether the agent invents, stalls, or escalates.
- Separate read access from write authority and name approval gates directly; “use judgment” is not an authorization rule.
- Test failure cases deliberately, including ambiguous requests, stale source material, and prohibited-action requests.
FAQ
What is the difference between an agent prompt and a normal prompt?
An agent prompt defines a repeatable operating procedure, input and output contracts, tool boundaries, and escalation behavior. A normal prompt can be shorter for a one-off task.
Should an AI agent be allowed to take actions automatically?
Only for bounded, reversible actions with clear authorization. Keep sends, purchases, account changes, legal commitments, and sensitive-data handling behind explicit approval rules.
How do I measure whether an agent is reliable?
Create a small evaluation set of real cases and edge cases. Score factual support, correct routing, output completeness, and compliance with the stated boundaries.