How to Write Agentic Prompts: A Six-Section Framework

Agentic prompts turn a broad goal into a controlled workflow that an AI can follow step by step. To write one, define six parts: role, objective, context, task, constraints, and output format. This prompt engineering framework creates a clear AI prompt structure for multi-step work.

Key takeaways

• Use an agentic prompt for work with connected steps, tools, evidence, or decisions.
• Assign a practical role with clear authority. Avoid vague labels such as “world-class expert.”
• Define one final outcome, the intended user, and testable success criteria.
• Provide trusted sources, current inputs, and rules for missing or conflicting information.
• Break the work into ordered steps with checks and approval points.
• Set strict limits for privacy, spending, publishing, communication, and account access.
• Keep a person in control of sensitive, costly, irreversible, or high-stakes actions.

What makes a prompt agentic?

An agentic prompt delegates a workflow instead of asking a single question. It gives the AI a goal, relevant information, permitted tools, ordered tasks, decision rules, boundaries, and a definition of success.

A simple prompt asks for an immediate answer:

Summarize this customer interview.

An agentic version assigns connected work:

Extract the customer’s problems. Group them by theme. Support each theme with a direct quote from the interview. Flag uncertain interpretations. Rank the themes by how often they appear, and produce a table for the product team.

The second version asks the AI to transform information, make limited judgments, check its work, and produce a useful deliverable.

When agentic delegation is useful

Use an agentic prompt when a task includes several of these elements:

• Research across approved sources
• Analysis followed by a recommendation
• Several documents, files, or data inputs
• Repeated work, such as reviewing a group of job listings
• Tool use, such as search, spreadsheets, or document retrieval
• Dependencies, where one step must finish before another begins
• Quality checks, citations, or evidence requirements
• A structured deliverable for a specific audience

Common uses include comparing competitors, screening job listings, building a study plan, auditing a content brief, and turning meeting notes into assigned actions.

A simple prompt is often better for a definition, quick rewrite, short calculation, or summary of one clear source. A long workflow can make a small task harder than needed.

Set the right level of autonomy

“Agentic” does not mean unlimited authority. State what the AI may decide, what requires approval, and what it must never do.

A useful authority ladder is:

1. Analyze only: Review information and report findings.
2. Recommend: Rank options and explain trade-offs.
3. Draft: Prepare materials without sending or publishing them.
4. Act with approval: Stop at a checkpoint and wait for permission.
5. Act within limits: Complete only defined, reversible actions.

Choose the lowest level that can complete the task. Keep human approval for payments, account changes, private data, formal submissions, external messages, and major medical, legal, financial, or employment decisions.

The six-section framework for agentic prompts

A reliable prompt engineering framework separates instructions into six sections. This AI prompt structure makes gaps easier to find and helps you reuse the workflow.

Here is a copy-paste template:

Role: Act as a [specific function] with relevant knowledge of [subject]. You may [permitted decision], but you must not [restricted action].

>

Objective: Produce [deliverable] for [audience] so they can [decision or action]. Success means [testable criteria].

>

Context: Use [inputs and approved sources]. Treat [source] as authoritative. If sources conflict, [conflict rule]. Label missing information instead of guessing.

>

Task:

1. Inspect the inputs.

2. Ask questions only if missing information would materially change the result.

3. Complete [analysis or transformation].

4. Verify [claims, dates, calculations, or citations].

5. Pause for approval before [sensitive action].

>

Constraints: Stay within [scope, time, or budget]. Protect [private information]. Do not invent facts, quotes, sources, or results. Do not [prohibited actions].

>

Output Format: Return [format] with [required sections or fields]. Include [citations, confidence labels, unknowns, or checklist].

The sections do not need equal detail. A basic formatting task may need one sentence for each. A research workflow may need detailed source, verification, and approval rules.

Sections 1 and 2: Set the role and objective

Write a functional role

“Act as a genius marketer” gives little useful direction. It does not identify the market, audience, deliverable, or authority level.

A practical role states:

• The work function
• The relevant area of knowledge
• The intended audience
• The decisions the AI may make
• The decisions reserved for a person

Weak:

You are a world-class career coach.

Stronger:

Act as a resume editor familiar with US technology hiring. Evaluate the resume against the supplied job description. You may recommend edits and draft replacement bullets. Do not add skills, employers, dates, credentials, or results that the candidate’s notes do not support.

The stronger version defines the task and reduces the risk of unsupported claims. Listings from the AI Job Portal can be used as inputs, but the role should still prohibit false statements and automatic applications.

Match the role to the task

Ask for the smallest useful set of skills. Avoid assigning several unrelated roles unless the workflow truly needs them.

For example:

Act as a research analyst who can compare product positioning, check source dates, and separate direct evidence from interpretation.

This is clearer than asking the AI to act as a strategist, writer, designer, economist, and data scientist at the same time. If the task needs different skills, divide it into stages and define the role for each stage.

Define authority with action verbs

Clear verbs reduce confusion:

• Extract: Copy or identify information without adding interpretation.
• Classify: Place items into defined categories.
• Assess: Compare evidence with supplied criteria.
• Recommend: Suggest an option while leaving the decision to the user.
• Draft: Prepare a deliverable without publishing or sending it.
• Execute: Complete a permitted action.
• Escalate: Stop and request human review when a rule is triggered.

Add a stop rule when needed:

Do not submit, publish, message, purchase, change an account, or delete anything. Prepare the final materials and wait for approval.

Define one final outcome

The objective is the finish line, not a loose list of activities. “Research competitors and think of ideas” describes activity. A stronger objective names the deliverable and its purpose:

Produce a source-backed comparison of five selected competitors for a product manager deciding which onboarding gaps to test next.

A clear objective answers four questions:

1. What will be delivered?
2. Who will use it?
3. What decision or action will it support?
4. How will success be judged?

Add testable success criteria

“Measurable” does not always mean numerical. It means a reviewer can check whether the result meets the requirement.

Useful criteria include:

• Every changing factual claim links to an approved source.
• Each recommendation is connected to evidence.
• Missing information is labeled “not found.”
• The result uses every required field.
• The language suits the named audience.
• No unsupported facts appear.
• A final checklist confirms whether each requirement passed.

Copy-paste example:

Role: Act as a project operations assistant. You may organize tasks, identify dependencies, and suggest owners. Do not assign work or change deadlines without approval.

>

Objective: Turn the meeting transcript into an action plan that the project lead can scan quickly. Success means every action has an owner or “owner needed,” a due date or “date needed,” and a reference to the supporting transcript passage.

Sections 3 and 4: Add context and build the workflow

Provide a source hierarchy

Context tells the AI what information to use and which material takes priority. Attach or paste the relevant inputs. Do not assume the system can access your files, subscriptions, or internal records.

List sources from highest to lowest priority:

1. Supplied policies, contracts, briefs, or assignment instructions
2. Primary sources, such as official documentation or original data
3. Approved secondary sources
4. Background material used only to generate ideas

Add a conflict rule:

If the current project brief conflicts with an older meeting note, follow the current project brief and flag the conflict.

For current research, ask the AI to record publication or update dates when shown. Tell it to label missing or inaccessible facts instead of filling gaps with plausible text.

Separate facts, assumptions, preferences, and unknowns

These inputs have different meanings:

• Facts: Confirmed information, such as an audience or deadline
• Assumptions: Working beliefs that may need testing
• Preferences: Desired tone, structure, or method
• Unknowns: Missing details that could affect the result

Example:

Facts: The audience is first-year college students. The exam is on May 18.

>

Assumption: The student can study for 45 minutes on weekdays. Confirm this before building the schedule.

>

Preference: Use short review sessions and weekly practice tests.

>

Unknown: The syllabus weighting has not been provided.

This prevents the AI from treating every input as equally certain.

Name tools and their limits

State which tools are available and how each may be used. Depending on the system, tools may include web search, document retrieval, spreadsheets, code execution, or resources in the AI Tool Lab.

For example:

Use web search only to find current official product pages and documentation. Use the spreadsheet to calculate totals. Do not cite a search snippet as evidence; open the source page. Do not sign in to accounts or bypass access controls.

Do not assume every AI system can browse, open links, edit files, or run code. Add a fallback rule:

If a required tool or file is unavailable, stop and list what is needed. Do not pretend the step was completed.

Turn the task into ordered actions

Write each step as a verb with a visible result:

1. Inspect the supplied inputs and list missing items.
2. Ask up to three blocking questions.
3. Create a short work plan.
4. Gather evidence from approved sources.
5. Complete the analysis using the stated criteria.
6. Check claims, calculations, links, and source dates.
7. Produce the requested deliverable.
8. Run the acceptance tests before returning it.

This reduces the risk that the AI will produce polished output before checking the evidence.

Add clarification rules

Without clear rules, an AI may guess too much or interrupt for minor details. Define when it should ask, assume, or stop.

Useful instructions include:

Ask a question only if the missing information could materially change the recommendation. Otherwise, continue with a clearly labeled assumption.

Ask no more than three questions at one time.

If more than three essential inputs are missing, stop and provide a missing-information checklist.

Do not assume dates, prices, legal requirements, product features, or personal facts.

Add checkpoints

A checkpoint is useful before a costly, sensitive, or hard-to-reverse step. It can also stop a workflow built on a faulty assumption.

Examples:

After selecting the research sources, show the list and wait for approval before analyzing them.

Draft the email, but do not send it.

Present the proposed database changes and a rollback plan. Wait for written approval before making any permitted change.

Stop if the estimated cost exceeds [LIMIT].

Reusable workflows can be accessed across devices through the Moyan AI app, but each run still needs suitable inputs, current source rules, and the right authority level.

Require a separate verification step

“Be accurate” is too vague. Name the checks that matter:

• Open cited pages and confirm that they support the claim.
• Check publication or update dates when shown.
• Recalculate totals.
• Compare names, dates, and figures with the source material.
• Mark conflicting evidence.
• Remove claims that cannot be supported.
• State what could not be verified.
• Confirm that links point to the cited material.

A useful instruction is:

Before finalizing, check each material claim against its source. In the final answer, flag any statement that relies on an assumption or could not be confirmed.

For calculations, ask the AI to show the formula and input values. For quoted text, require the exact wording from the supplied source and enough location detail for a reviewer to find it.

Sections 5 and 6: Set constraints and control the output

Constraints define where the AI must stop. Output rules define what finished work should look like. Together, they make the result easier to review and reuse.

Define hard limits

Cover only the limits that matter for the task:

Use exact boundaries where possible. “Keep costs low” is vague. This is testable:

Use free sources only. Request approval before any action that could create a charge.

Separate hard constraints from preferences:

• Hard: Do not include private customer data.
• Hard: Do not submit a form or send an email.
• Preference: Keep the brief under 1,200 words.
• Preference: Favor recently updated sources when they are available.

If rules could conflict, state the priority. A practical order is:

1. Safety and privacy
2. Accuracy and source requirements
3. Scope and approval rules
4. Task completion
5. Speed and style preferences

Protect private information

Do not provide passwords, authentication codes, private keys, or confidential records unless the system and workflow are approved for that data. Remove identifiers that are not needed. Replace them with labels such as [CLIENT A] or [ORDER ID].

Add a privacy rule:

Use only the supplied redacted files. Do not infer missing personal details. Do not reproduce email addresses, phone numbers, or account identifiers in the output. If sensitive information appears unexpectedly, stop and flag it.

For workplace tasks, follow the organization’s data, access, and retention policies. For medical, legal, tax, or investment topics, use AI to summarize supplied information, organize records, and prepare questions for a qualified professional. Do not let it present itself as that professional.

Require traceable research

“Use reliable sources” is too broad. Define acceptable evidence for the specific task.

A research rule might say:

• Prefer primary sources, such as official documentation, public filings, original research, and government pages.
• Use reputable secondary reporting when primary evidence is unavailable.
• Link factual claims that could have changed.
• Record publication or update dates when shown.
• Mark unsupported claims as “not verified.”
• Never create a citation, quotation, publication date, or source detail.
• Do not treat search snippets or AI-generated summaries as final evidence.

For time-sensitive work, add:

Verify availability and other changing details on the day the research is performed. State the verification date in the result.

Name prohibited actions

Common prohibited actions include:

• Sending emails, messages, applications, or forms
• Publishing or deleting content
• Buying products or starting subscriptions
• Accepting terms on the user’s behalf
• Changing a live database
• Contacting research participants
• Claiming credentials or experience the user does not have
• Bypassing access controls, paywalls, or website restrictions
• Uploading private data to an unapproved service

A safe pattern is draft, review, execute. The AI prepares the action. A person reviews it. An authorized person then completes it or approves it within defined limits.

Specify the output contract

An output contract defines the exact deliverable. It can include headings, fields, tables, file types, length limits, and labels.

For a research brief, require:

1. Executive summary
2. Findings table
3. Evidence and source links
4. Risks and unknowns
5. Recommended next actions

For data passed into another tool, use consistent field names:

• company_name
• finding
• evidence_url
• source_date
• confidence
• needs_review

Request strict JSON only when another system needs it. For human review, headings and a Markdown table are often easier to inspect.

Add acceptance tests

Acceptance tests are pass-or-fail checks completed before delivery.

Example:

Before delivery, confirm that each selected company has the required evidence, every link points to the cited page, no unsupported market-size figure appears, all unknowns are labeled, and each recommendation follows from the evidence. If a test fails, revise the work or list the failure.

A compact quality report can include:

Do not rely on self-checking alone for important work. A person should still inspect key evidence, calculations, and decisions.

Complete agentic prompt examples

Each example uses the same six sections. Replace the bracketed text with your own details.

Market research prompt

Role: Act as a market research analyst. Compare conflicting claims and separate facts, interpretations, and unknowns. You may recommend a next step, but you may not make purchases or contact companies.

>

Objective: Produce a decision brief on whether [PRODUCT] should target [CUSTOMER GROUP] in [REGION]. Success means the brief identifies customer needs, relevant competitors, likely barriers, supporting evidence, and unanswered questions.

>

Context: The product is [DESCRIPTION]. The decision is needed by [DATE]. Use the attached notes and current public sources. Prefer official or primary sources.

>

Task:

1. Ask up to three questions if missing details could change the research direction.

2. Define the proposed market and customer segment.

3. Find evidence of customer needs, current alternatives, and recurring complaints.

4. Compare up to five selected competitors using the same criteria.

5. Check major claims against primary sources when available.

6. Separate confirmed findings, reasonable inferences, and unknowns.

7. Recommend proceeding, pausing, or running a limited test.

>

Constraints: Use public sources that can be linked. Do not invent market-size figures, quotes, or citations. Do not purchase reports or contact companies. Stay within [TIME LIMIT]. Treat the result as business research, not investment advice.

>

Output Format: Return a short summary, a competitor table, five key findings with citations, risks, unknowns, and an example 30-day validation plan. End with an acceptance checklist.

This version limits the recommendation, requires traceable sources, and gives uncertainty a clear place in the output.

Job application prompt

Save suitable listings from the AI Job Portal, then review one listing at a time.

Role: Act as a careful job application editor. Preserve the applicant’s true experience and natural writing style.

>

Objective: Tailor the supplied resume and create a cover letter for [JOB TITLE] at [COMPANY] without overstating qualifications.

>

Context: Inputs are the full job listing, current resume, and a list of verified achievements. The audience is a recruiter and hiring manager in [COUNTRY].

>

Task:

1. Extract the role’s main requirements.

2. Map each requirement to verified evidence from the resume and notes.

3. Identify gaps without trying to hide them.

4. Rewrite the summary and selected bullets using relevant terms naturally.

5. Draft a cover letter.

6. Prepare five interview questions based on the role and identified gaps.

>

Constraints: Do not invent skills, dates, employers, degrees, metrics, credentials, or certifications. Do not submit the application. Do not add demographic details. Use simple headings and formatting unless another format is requested. Flag every statement that needs verification.

>

Output Format: Provide a requirement-to-evidence table, revised resume sections, a cover letter of no more than 350 words, identified gaps, and five interview questions. End with “Facts to verify before submission.”

The AI may analyze and draft, but it may not fabricate or apply without approval.

Study planning prompt

Role: Act as a study planner and tutor for [SUBJECT] at [LEVEL].

>

Objective: Build an example four-week plan for an exam on [DATE], with more time assigned to weak and important topics.

>

Context: Available study time is [MINUTES] on weekdays and [MINUTES] on weekends. The current scores and syllabus are attached. Preferred methods are practice questions, active recall, and short reviews.

>

Task:

1. Identify weak areas from the supplied results.

2. Rank topics by syllabus importance and current weakness.

3. Schedule realistic study blocks.

4. Include weekly checks.

5. Give rules for adjusting the following week based on results.

>

Constraints: Use only the attached syllabus to identify examinable topics. Keep each session under [LENGTH]. Include one rest day each week. Do not promise a grade.

>

Output Format: Give a weekly table with date, topic, activity, duration, and completion box. Add a checkpoint quiz plan, adjustment rules, and required materials.

The workflow adapts to supplied evidence instead of dividing time equally without a reason.

Content production prompt

Writing, summarizing, or document tools from the AI Tool Lab may support relevant stages.

Role: Act as a content editor for a practical publication serving [AUDIENCE].

>

Objective: Create an accurate article that answers [SEARCH QUERY] and helps readers complete [ACTION].

>

Context: Use the content brief, approved sources, brand guide, and internal-link list. Readers have beginner knowledge but want specific instructions.

>

Task:

1. Identify the reader’s main need.

2. Build an outline.

3. Research claims using approved sources.

4. Draft the article.

5. Check every factual statement that could have changed.

6. Improve headings, examples, and transitions.

7. Remove repetition and unclear wording.

>

Constraints: Do not invent studies, statistics, quotes, dates, sources, or product capabilities. Cite time-sensitive claims. Use plain English and short paragraphs. Do not publish. Request approval if the brief conflicts with an approved source.

>

Output Format: Return the outline, finished Markdown article, source log, claims needing review, and a checklist covering accuracy, intent, readability, and links.

Research, drafting, and verification are separate steps, which makes problems easier to find.

How to test and improve agentic prompts

Test each workflow with a realistic, low-risk example before granting more authority. Review the output and any visible record of tool use, sources, assumptions, or failed steps.

Use an evaluation checklist

• Goal: Is the desired outcome clear and testable?
• Inputs: Are all required files, dates, definitions, and source rules present?
• Authority: Does the AI know what it may decide or do?
• Questions: Does it ask only when missing information matters?
• Evidence: Can a reviewer trace important factual claims?
• Tools: Are approved tools and prohibited actions clear?
• Privacy: Has sensitive information been removed or protected?
• Output: Does the result match the required structure?
• Quality: Do acceptance tests catch missing or weak work?
• Recovery: Does the AI report failed steps instead of hiding them?
• Approval: Does it stop before major actions?

Fix common failures

Use a short improvement loop

1. Test: Run one realistic, low-risk task.
2. Inspect: Mark unsupported claims, missed instructions, and unusable output.
3. Diagnose: Find the likely cause. It may be missing context, unclear authority, weak constraints, or a poor output contract.
4. Edit: Change the smallest section that can fix the problem.
5. Retest: Use the same input, then try a different example.
6. Version: Save the template with a purpose, revision date, and short change note.

You can store reusable templates through a free Moyan AI account if it fits your workflow. Review what Moyan AI includes before choosing where to use each template. For access across devices, install the Moyan AI app.

Set a stopping rule

Research and revision can continue forever unless the prompt defines “enough.” Add a rule such as:

Stop research when each required question has at least one suitable primary source or two suitable independent secondary sources, or when the time limit is reached. Report any question that remains unresolved.

The right stopping rule depends on the task. A quick content brief may use a time limit. A compliance review may require every checklist item to be resolved or sent for human review.

Plan for failure

Tool calls, websites, files, and calculations can fail. Tell the AI how to recover:

If a source is unavailable, do not replace it with an unsupported claim. Record the failed source, try one approved alternative, and label the gap if it remains unresolved.

If a calculation produces an unexpected result, check the inputs and formula once. If the issue remains, stop and show the values used.

If instructions conflict, follow the stated priority order and report the conflict.

A failure report is more useful than a confident but incomplete answer.

Frequently asked questions

What is the difference between prompt engineering and agentic prompting?

Prompt engineering is the broad practice of designing instructions and context for an AI system. Agentic prompting focuses on multi-step work that may include tools, decisions, verification, checkpoints, and limited autonomy.

How long should an agentic prompt be?

It should be long enough to remove important ambiguity, but no longer. A routine task may need six short sections. Research-heavy or sensitive work usually needs more detailed source, privacy, authority, and verification rules.

Delete instructions that do not change the result. Keep details that affect evidence, safety, scope, decisions, or the final deliverable.

Should an agentic prompt always use all six sections?

Consider all six for meaningful multi-step work. Some sections can be brief, but missing context, constraints, or output requirements often causes avoidable errors.

For a small task, combine sections if that makes the instructions clearer. The labels matter less than covering the required decisions.

When should the AI ask a clarifying question?

It should ask when missing information could materially change the result, break a hard constraint, create risk, or waste substantial work. For minor gaps, it can use a clearly labeled assumption if the instructions allow it.

Set a question limit and a stop rule. This prevents both excessive guessing and endless questioning.

Can an agentic prompt make an AI fully autonomous?

A prompt can grant limited autonomy within defined boundaries. It does not remove the need for human review.

Keep people in control of publishing, spending, external communication, account access, private information, formal submissions, and other major actions. Start with analysis or drafting, test the workflow, and expand authority only when there is a clear need.

Build your first six-section template

Choose one recurring, low-risk task. Write its Role, Objective, Context, Task, Constraints, and Output Format.

Then test it with this quick checklist:

• Is the final outcome clear?
• Are the inputs and trusted sources named?
• Are actions ordered?
• Are missing-information rules included?
• Is the authority level limited?
• Are private or prohibited actions covered?
• Is the deliverable easy to inspect?
• Are there pass-or-fail acceptance tests?

Revise the smallest weak section, test again, and keep human approval at every major checkpoint.