How to Write AI Agent Instructions With Real Examples
Quick Answer
To write effective AI agent instructions, define one job, the available context, the process, and the limits.
Tell the agent what a good result looks like.
Add examples for tasks where consistency matters.
Then test the instructions using realistic inputs and edge cases.
AI Summary
Strong AI agent instructions read like a clear operational brief. They specify the agent’s purpose, information sources, decision rules, output format, and escalation path. The best instructions do not try to predict every possible situation. Instead, they set useful defaults and state when the agent must pause, ask, or hand work to a person.
What This Guide Covers
- The difference between a prompt and an agent instruction set
- A six-part framework for reliable agent behaviour
- Real instruction examples for research, client communication, and reporting
- Common mistakes that create vague or risky outputs
- A testing method for improving an agent over time
What Makes AI Agent Instructions Work?
Effective instructions reduce ambiguity before the agent starts working. They give it a clear destination, helpful constraints, and a dependable way to handle uncertainty.
A chat prompt may be one request. Agent instructions guide repeated work across many requests, documents, tools, and decisions. That makes them closer to a job description and process manual than a one-off question.
Anthropic’s guidance on effective agents makes a useful distinction: use the simplest pattern that can complete the work reliably. Not every task needs an autonomous agent. Some tasks work better as a structured workflow with defined steps.
| Instruction Element | What It Answers | Example |
|---|---|---|
| Role | Who is the agent? | “You are a research assistant for a financial advisory firm.” |
| Goal | What must it achieve? | “Create a source-backed briefing for the adviser.” |
| Context | What can it use? | “Use the attached client notes and approved research sources.” |
| Workflow | How should it work? | “Extract facts, identify gaps, then draft a concise summary.” |
| Boundaries | What must it avoid? | “Do not make suitability recommendations or send client emails.” |
| Output | What does success look like? | “Use headings, source links, risks, and open questions.” |
Clear instructions also improve review. A manager can see what the agent was intended to do, compare that with its output, and update one rule at a time.
How Do You Write AI Agent Instructions Step by Step?
The best way to write AI agent instructions is to move from job definition to controlled execution. Start simple, then add detail only where real work requires it.
1. Define One Specific Job
Avoid giving one agent every responsibility in a department. A broad instruction such as “help with client service” invites inconsistent decisions.
Instead, define a narrow, observable outcome:
Create a first draft of a meeting brief using the client’s latest notes, portfolio documents, and agenda.
This makes testing simpler. It also clarifies which information the agent needs.
OpenAI’s developer guidance recommends clearly defining agent responsibilities and the tools available for a task. The same principle applies whether your team is coding an agent or configuring one through a no-code platform.
2. State What the Agent Can Use
An agent should not guess which data is relevant. Name the approved sources, expected inputs, and any source hierarchy.
For example:
Use the meeting transcript, the client profile, and the approved planning template. If sources conflict, flag the conflict instead of deciding which source is correct.
This simple rule prevents a common failure: the agent blending old and new information without telling the reviewer.
3. Describe the Working Process
You do not need to prescribe every thought. However, you should specify the checkpoints that matter.
A practical workflow might be:
- Read the input documents.
- Extract relevant facts.
- Identify missing information.
- Produce the required draft.
- Flag assumptions and unresolved issues.
Google’s prompt design guidance similarly stresses clear task instructions, formatting requirements, and examples when you need more dependable outputs.
4. Add Boundaries and Escalation Rules
Boundaries are not negative instructions added at the end. They are part of the job.
Include rules such as:
- Do not invent facts, citations, or client details.
- Do not take irreversible actions without approval.
- Ask a clarifying question if required information is missing.
- Escalate legal, regulatory, medical, or financial advice requests.
- State uncertainty clearly when evidence is incomplete.
These rules help the agent choose a safe next step rather than producing a confident but unreliable answer.
5. Specify the Output Format
“Write a good report” is subjective. A structured result is easier to review and reuse.
For a client meeting brief, request:
- A five-bullet summary
- Key decisions from the previous meeting
- Open questions for the adviser
- Relevant risks or missing documents
- A draft agenda
Microsoft’s guidance on prompt engineering supports this approach. Clear structure reduces interpretation and gives users more predictable outputs.
6. Provide Examples, Then Test Them
Examples show the level of detail, tone, and judgement you expect. They are especially valuable when the agent must classify, summarise, or draft in a consistent house style.
The goal is not to add dozens of examples. Use two or three representative cases, including one edge case.
What Is a Reusable AI Agent Instruction Template?
A reusable template gives teams a consistent starting point. It also makes agents easier to review, maintain, and share.
ROLE
You are [agent role] for [team or business].
PRIMARY GOAL
Your job is to [specific outcome].
WHEN TO ACT
Start when [trigger or user request].
APPROVED INPUTS
Use only [documents, systems, tools, or user-provided information].
PROCESS
1. [Step one]
2. [Step two]
3. [Step three]
DECISION RULES
- If [condition], then [action].
- If information is missing, [clarify or flag it].
- If sources conflict, [state the conflict].
BOUNDARIES
- Do not [prohibited action].
- Do not invent information.
- Do not take [sensitive action] without human approval.
OUTPUT FORMAT
Provide:
1. [Section or field]
2. [Section or field]
3. [Section or field]
QUALITY CHECK
Before responding, confirm that you followed the output format and flagged uncertainty.
This template works because it separates purpose from process. The agent can understand its job before it encounters detailed rules.
How to Write AI Agent Instructions With Real Examples
The following examples show how a vague prompt becomes a usable instruction set.
Example 1: Research Briefing Agent
Weak instruction
Research this company and tell me what matters.
The agent has no audience, source standard, or expected output. It may return a long summary without practical next steps.
Improved instruction
You are a business research assistant.
Create a concise company briefing for a consultant preparing for a first meeting.
Use only the company website, official filings, and reputable primary sources.
Do not treat marketing claims as verified facts.
Follow this process:
1. Identify the company’s products, customers, leadership, and recent developments.
2. Separate verified facts from reasonable inferences.
3. Flag missing information and source conflicts.
Output:
- Company overview, maximum 100 words
- Three relevant recent developments
- Likely business priorities
- Five meeting questions
- Sources as direct links
If a source is unavailable, say so. Do not fill gaps with guesses.
This agent has a role, an audience, a source standard, a process, and a usable output.
For teams using connected tools, tool access needs equally clear rules. LangChain’s overview of tool calling explains why tool use should be explicit and structured rather than assumed.
Example 2: Client Email Drafting Agent
Weak instruction
Reply professionally to clients.
This fails because “professionally” means different things to different people. It also offers no rule for sensitive requests.
Improved instruction
You are a client communications drafting assistant.
Draft clear, warm email replies for a professional services firm.
Your drafts are always reviewed by a team member before sending.
Use the client’s name, the message context, and the firm’s approved tone guidance.
Keep emails under 180 words unless the user asks for more detail.
Do:
- Confirm the client’s request.
- State the next step and owner.
- Ask one clear question if information is missing.
Do not:
- Give legal, investment, tax, or compliance advice.
- Promise a deadline that is not stated in the source material.
- Send, schedule, or finalise communications.
Output only the draft email and a short “Review Notes” section.
The key improvement is not the wording. It is the operating boundary. The agent drafts. A person approves.
On LaunchLemonade’s Teams platform, admins can define which agent actions need human review before they run. That is useful for sensitive tasks, such as client emails or final reports.
Example 3: Monthly Reporting Agent
Weak instruction
Create our monthly report.
The agent needs to know which numbers matter, how to treat gaps, and who will read the result.
Improved instruction
You are an operations reporting assistant.
Prepare a monthly performance report for the leadership team.
Use the approved spreadsheet and meeting notes only.
Check that all figures include a reporting period and unit.
Process:
1. Summarise the five agreed performance metrics.
2. Compare each metric with the previous month.
3. Identify material changes.
4. List unresolved data-quality issues separately.
Output:
- Executive summary, maximum 150 words
- Metrics table
- Key changes and likely causes
- Risks and decisions needed
- Data gaps
Do not calculate figures from incomplete data.
If a value is missing, write “Data unavailable” and identify the owner.
This agent creates a decision-ready report while protecting against silent assumptions.
Why Do Agent Instructions Need Guardrails?
Guardrails tell an agent when not to continue. They are essential when the work affects clients, compliance, data, or external systems.
A useful guardrail is specific and actionable. “Be careful with data” is not actionable. “Do not use client data outside the approved workspace” is clearer.
| Risk Area | Weak Guardrail | Strong Guardrail |
|---|---|---|
| Missing data | “Use your judgement” | “Ask for the missing document before drafting conclusions.” |
| Sensitive action | “Be cautious” | “Prepare a draft only. Human approval is required before sending.” |
| Source quality | “Use reliable sources” | “Use primary sources first. Label unverified claims as unverified.” |
| Advice risk | “Do not give advice” | “Provide general information only and escalate personalised recommendations.” |
| Tone | “Be professional” | “Use plain English, short paragraphs, and no promotional language.” |
The NIST AI Risk Management Framework is useful background for teams formalising controls. Its core message is practical: manage risk throughout the AI lifecycle, not only after an incident.
LaunchLemonade is built for regulated small and medium businesses that need governance around AI agents. Teams can customise ready-made agents with their templates, tone, source documents, and workflows, or build their own through the no-code builder. Explore the Teams platform to see how role-based access controls, audit trails, PII detection, and approval workflows can support governed agent use.
What Mistakes Make Agent Instructions Fail?
Most instruction failures come from ambiguity, not from a lack of clever prompting.
Trying to Solve Everything With One Agent
A single agent that researches, drafts, approves, sends, and reports creates too many competing rules. Split the work into separate agents or a defined workflow.
This aligns with Anthropic’s workflow and agent guidance: begin with the simplest reliable structure, then add complexity only when the task needs it.
Hiding Critical Rules in Long Paragraphs
Long narrative instructions are easy to write and difficult to maintain. Use headings, lists, and decision rules so a reviewer can identify the important constraints quickly.
Leaving “Good” Undefined
Words such as accurate, concise, professional, and useful need evidence. Define the source standard, word limit, headings, tone rules, and expected audience.
Forgetting Failure Behaviour
Your agent needs a response for incomplete, conflicting, or sensitive inputs. “Ask”, “flag”, “stop”, and “escalate” are often more valuable than more detail about the happy path.
Never Testing With Real Inputs
A polished example can hide failures. Test instructions against routine cases, difficult cases, missing data, and conflicting directions.
OpenAI’s evaluation guidance is helpful here. Reliable AI work improves through defined test cases and repeatable evaluation, not one successful demo.
How Should You Test AI Agent Instructions?
Test the agent as you would test a new team process. Use examples drawn from real work, remove sensitive information where needed, and decide what “pass” means before reviewing outputs.
Build a Small Test Set
Start with 10 to 20 realistic tasks:
- A standard request
- A request with missing information
- Conflicting source documents
- An ambiguous request
- A sensitive request that requires escalation
- A request outside the agent’s scope
Score the Results
| Test Question | Pass Standard |
|---|---|
| Did the agent complete the assigned task? | The result addresses the requested outcome. |
| Did it use the approved inputs? | No unsupported claims or irrelevant sources appear. |
| Did it follow the output format? | All required headings and fields are present. |
| Did it handle uncertainty correctly? | Gaps and conflicts are clearly flagged. |
| Did it respect boundaries? | It did not take prohibited or unapproved actions. |
Google’s evaluation guidance recommends using systematic evaluation rather than relying on ad hoc impressions. That is particularly important when multiple people will depend on the agent’s output.
Track failures by category. For example, “missed source conflict” is more useful than “bad answer.” Update the smallest rule likely to resolve that category, then rerun the same test set.
How Should Teams Govern Agent Instructions?
Teams should treat agent instructions as controlled operational assets. Someone should own each agent, approve material changes, and review performance after workflows or source documents change.
LaunchLemonade makes this easier for non-technical teams. Its no-code builder lets domain experts create and update agents without engineering support. You can also create structured, multi-step workflows with decision points, tool calls, output formatting, schedules, or event triggers.
For consultants and professional-services experts who want to package specialist workflows, LaunchLemonade’s Builders platform is designed for creating no-code agents around a clear client problem.
A practical governance process has five parts:
- Named owner: One person owns the task definition and quality standard.
- Version control: Record what changed and why.
- Approval rules: Require review for high-impact actions.
- Access control: Limit data and tools to what each agent needs.
- Periodic tests: Re-run key scenarios after model, workflow, or policy changes.
The UK Information Commissioner’s Office guidance on AI is also relevant for organisations handling personal data. Data protection obligations do not disappear because an AI system performs part of the work.
Key Takeaways
- Effective agent instructions define one job, not an entire department.
- Specify approved inputs, workflow steps, outputs, and boundaries.
- Use examples when tone, classification, or judgement must stay consistent.
- Tell the agent what to do when information is missing or conflicting.
- Test with realistic and difficult cases, not only ideal examples.
- Treat instructions as living operational documents with named ownership.
Conclusion
Knowing how to write AI agent instructions is less about finding magic wording. It is about giving an agent the same clarity you would give a capable new colleague.
Start with one narrow workflow. Define the expected result. Set clear limits. Then test, revise, and govern the agent as the work evolves.
If your firm needs agents that can work with real operational controls, book a LaunchLemonade demo. The team can show how governed agents, workflows, approvals, and access controls fit your existing processes.
Frequently Asked Questions
What Should AI Agent Instructions Include?
AI agent instructions should include a role, goal, approved inputs, process, boundaries, output format, and escalation rules. Include examples when consistent judgement or formatting matters.
How Long Should AI Agent Instructions Be?
They should be as short as possible while remaining unambiguous. Clear rules and useful examples matter more than total length.
Should an AI Agent Have One Task or Many?
Start with one well-defined job. Separate agents or structured workflows usually produce clearer, easier-to-test results.
What Is the Difference Between a Prompt and Agent Instructions?
A prompt is usually a one-time request. Agent instructions define repeated behaviour across many tasks, inputs, and decisions.
How Do I Stop an Agent From Making Things Up?
Tell it to use only approved sources, flag gaps, and state uncertainty. Test this rule with missing and conflicting information.
When Should an AI Agent Ask for Human Approval?
It should ask before taking high-impact or irreversible actions. Examples include sending client communications, finalising reports, or updating connected systems.
Do AI Agent Instructions Need Examples?
Examples are helpful when an agent must follow a particular tone, format, classification method, or quality bar. Use a few representative examples rather than many repetitive ones.
How Often Should We Review Agent Instructions?
Review them after meaningful workflow, policy, tool, or source-document changes. Also review them when test results reveal recurring failure patterns.
