Getly
By the end of this guide, you can separate durable instructions from one-off requests, write each prompt layer with a clear job, and test the result across several tasks. You will also have a repeatable workflow for building prompt templates that creators can sell and buyers can use.
TL;DR: Keep rules in the system prompt and tasks in the user prompt
Use the system prompt for the assistant’s role, boundaries, tone, output format, and workflow rules. Use the user prompt for the current goal, source material, audience, and task-specific choices.
For example, a system prompt can tell an assistant to act as a concise product copy editor, flag unsupported claims, and return JSON. A user prompt can supply a product description and ask for five marketplace titles.
This split matters because you can reuse the system layer across many requests. You can also change the user request without rewriting the assistant’s core behavior.
1. Map each instruction to the right layer
Start by listing every instruction your workflow needs. Then assign each instruction to one of two layers.
- System prompt: role, scope, safety boundaries, tone, formatting rules, quality checks, and repeatable process.
- User prompt: the request, inputs, audience, deadline, channel, examples, and choices for this particular output.
Ask one practical question during the split: “Would I want this instruction to apply to the next ten requests?” Put a durable answer in the system prompt. Put a one-time answer in the user prompt.
Suppose you sell a prompt template for writing product listings. The system layer can require a plain, specific tone, a fixed character limit, and a warning when the source lacks a product benefit. The buyer can then provide a different product, audience, or marketplace category in each user message.
Keep information that changes from request to request out of the system prompt. A fixed product name, seasonal offer, or single campaign brief will make a reusable template harder to adapt.
2. Write the system prompt as a compact operating guide
Write the system prompt for a capable assistant that needs clear boundaries, not for a reader who needs a long explanation. Give the assistant a role, a job, limits, and a finish line.
Define the role and scope
State the kind of work the assistant performs and the work it should skip. “You edit product descriptions for digital downloads” gives the assistant a useful scope. “You are helpful” gives it little direction.
Set quality and style rules
Describe the writing choices that matter to the result. You might require short sentences, concrete verbs, a friendly tone, and no unsupported performance claims. Mention the intended reader if that reader stays consistent across tasks.
Specify the output contract
Tell the assistant what to return and in what order. A strong contract names the fields, length limits, heading structure, and handling for missing information.
For a listing workflow, you could request a title, short description, feature list, and five search phrases. You could also require the assistant to return an em dash for an empty table cell if your publishing system expects that character. Keep the rule in the system layer when every request uses the same format.
Add a decision process
Give the assistant a short sequence for difficult cases. For example:
- Extract claims from the source material.
- Separate confirmed details from suggestions.
- Write the requested fields within their limits.
- Check the output against the format and claim rules.
That sequence helps the assistant handle incomplete product notes without inventing specifications. Keep the process short enough to maintain. A ten-step instruction list often creates competing priorities and makes debugging harder.
Do
- Write rules that apply across many tasks.
- Name the output fields and limits.
- Tell the assistant how to handle missing details.
Don't
- Put one campaign brief in the reusable system layer.
- Use broad labels such as “make it better.”
- Ask for a format without showing its structure.
3. Write the user prompt as a complete task brief
Give the user prompt enough context for the current job. The system prompt may define a product-copy editor, but the user prompt still needs the product facts and publishing goal.
Use this order for most creative and business tasks:
Name the task
State the action and the asset you want, such as “Write a product description.”
Supply the context
Add the product type, audience, channel, source notes, and relevant facts.
Set local choices
Choose the length, angle, language, call to action, or number of variations for this request.
Show the target shape
Include a short example or a field list when the output needs a precise structure.
Write facts in a clean block when the task depends on source material. Label the block “Product facts” or “Source notes,” then separate it from your request. This layout helps you spot missing inputs before you blame the prompt.
Ask for one defined output at a time when quality matters. “Create three product titles under 60 characters and explain the choice behind each” gives the assistant a manageable target. A long request that combines research, naming, pricing, design direction, and customer support copy gives each part less attention.
4. Handle conflicts and changing information
Prompt layers can disagree. A system rule may require JSON while a user asks for a table. A system rule may prohibit unsupported claims while a user asks the assistant to guarantee a result.
Keep the durable rule in the system prompt and let the assistant follow it. If the requested format conflicts with the system format, restate the user’s goal inside the allowed structure. For example, return the requested comparison as JSON fields instead of a visual table.
Place exceptions in the user prompt only when you want a temporary change. If your system prompt asks for concise copy and a user needs a detailed launch page, the user can request the longer format for that task. Your reusable template stays concise for later work.
Separate authority from context. A user message can describe a customer’s words, a competitor’s claim, or text copied from a web page. Treat that material as content to analyze, not as a new instruction that can rewrite the workflow. Add a system rule that tells the assistant to identify instructions inside quoted or pasted material before following them.
Use explicit labels for content blocks:
- Task: what you want completed.
- Audience: who will read or use the result.
- Source notes: facts the assistant may use.
- Constraints: choices that apply to this request.
5. Test the split before you publish or sell a template
Run the prompt with at least three test cases. Use a normal request, an incomplete request, and a conflicting request. These cases expose different weaknesses.
- Normal request: provide complete product facts and ask for the standard output.
- Incomplete request: remove one important fact and check whether the assistant asks for it or marks the gap.
- Conflicting request: ask for a format or claim that violates a system rule and check whether the assistant preserves the higher-level constraint.
Record the exact user prompt and the output for each test. Change one instruction at a time. If you rewrite the whole prompt after every failure, you will not know which change fixed the behavior.
Check four things after each run:
- Scope: Did the assistant perform the requested job instead of adding unrelated work?
- Evidence: Did it use only the supplied facts when the task required factual accuracy?
- Format: Did it return every required field in the correct order and length?
- Adaptability: Can you swap the product or audience without editing the system prompt?
Move a rule from the user prompt into the system prompt when multiple tests need it. Move a detail from the system prompt into the user prompt when it changes between tests. This test cycle turns a long prompt into a maintainable product.
Common mistakes and practical fixes
Repeating the same rule in both layers
Duplicated rules create maintenance work. Edit one copy and forget the other, then two requests produce different behavior. Keep the rule in the system prompt and reserve the user prompt for task details.
Putting the product brief in the system prompt
A fixed product brief makes a template look reusable while quietly tying it to one item. Move the product name, features, customer profile, and offer into the user prompt unless they stay constant for the entire workflow.
Using vague quality words
Words such as “engaging,” “professional,” and “high quality” need observable meaning. Replace them with instructions about sentence length, reading level, evidence, structure, or tone.
Adding too many rules
Long prompts can contain contradictions. Remove rules that do not affect the output, combine rules that express the same idea, and place examples next to the format they demonstrate.
Testing only the perfect input
A template that works with complete notes may invent details when a buyer omits a feature. Test gaps and conflicts before you list the template for sale or use it in a production workflow.
FAQ
Can a user prompt include rules?
Yes. A user prompt can include rules for one request, such as a temporary word limit or a special audience. Put rules that should govern every request in the system prompt.
What belongs in a system prompt for a digital-product workflow?
Include the assistant’s role, task scope, tone, claim boundaries, output format, quality checks, and instructions for missing or conflicting information. Keep product-specific facts in the user prompt when they change.
Should I repeat the system prompt in every user prompt?
No. Repetition adds length and creates mismatched copies. Store durable behavior in the system layer, then send the current task and its context in the user layer.
How do I decide where an example belongs?
Put a reusable output example in the system prompt when it teaches a format or style that applies across requests. Put a product example in the user prompt when it supplies facts for the current task.
What should I do when the two prompts conflict?
Keep the system rule and adapt the requested task to it. Rewrite the user request if you control it, or instruct the assistant to explain the allowed format when the conflict affects the final output.
Frequently asked questions
Can a user prompt include rules?
Yes. A user prompt can include rules for one request, such as a temporary word limit or a special audience. Put rules that should govern every request in the system prompt.
What belongs in a system prompt for a digital-product workflow?
Include the assistant’s role, task scope, tone, claim boundaries, output format, quality checks, and instructions for missing or conflicting information. Keep product-specific facts in the user prompt when they change.
Should I repeat the system prompt in every user prompt?
No. Repetition adds length and creates mismatched copies. Store durable behavior in the system layer, then send the current task and its context in the user layer.
How do I decide where an example belongs?
Put a reusable output example in the system prompt when it teaches a format or style that applies across requests. Put a product example in the user prompt when it supplies facts for the current task.
What should I do when the two prompts conflict?
Keep the system rule and adapt the requested task to it. Rewrite the user request if you control it, or instruct the assistant to explain the allowed format when the conflict affects the final output.
