KoreaDevKNOWLEDGE SHARING

Content typeLearn

PROMPT EDUCATION · 01 / 1

When you describe the task A usable prompt

Choose a task from development, questions, slides, writing, analysis, research, meetings, or translation. Answer the questions in order to create a final prompt with a goal, context, success criteria, constraints, and output format.

Difficulty
Absolute beginner
Structure
Lessons 5 · Labs 2 · Assessment

CORE UNIT 1 / 1

When you describe the task A usable prompt

Choose a task from development, questions, slides, writing, analysis, research, meetings, or translation. Answer the questions in order to create a final prompt with a goal, context, success criteria, constraints, and output format.

Difficulty
Absolute beginner
Structure
Lessons 5 · Labs 2 · Assessment

NEW HIRE ONBOARDING

Start in the order you would receive your first assignment

So that even a new hire with no prior IT background can follow along, we start with the situation, the task, the evidence, and when to report, before difficult definitions.

  1. 01

    Read the situation in one sentence

    Choose a task from development, questions, slides, writing, analysis, research, meetings, or translation. Answer the questions in order to create a final prompt with a goal, context, success criteria, constraints, and output format.

  2. 02

    Today's assignment

    Multilingual final prompt

  3. 03

    Evidence that shows the work is complete

    Is there a condition the reader can use to confirm that the answer is complete?

  4. 04

    When to stop and ask a senior colleague

    Replace passwords, token values, national identification numbers, original customer text, and private source code with pseudonyms, categories, or the minimum necessary excerpts. Review the generated prompt again before sending it.

Unpack unfamiliar terms first

Start with the outcome
Instead of simply writing “analyze this,” specify the decision to support and the deliverable you need.
Provide facts and context
Provide internal terms the model cannot know, the current state, the reference date, and source text as clearly separated inputs.
Write the pass line
List items that confirm completion and the conditions for failure, rather than aiming for an answer that merely looks good.

Input is processed only in this browser and is never sent to a server or AI.

STEP 01 · CHOOSE THE WORK

What kind of work do you do?

After you choose a task, only the necessary questions are asked, in order.

01

Development and debugging

Feature implementation, bug fixes, code review, and test planning

Implementation, verification, and change summary
02

Questions · learning

Concept understanding, technical questions, tutoring, and exam preparation

Understandable explanations and comprehension questions
03

Slide decks

Presentations for reporting, lectures, proposals, and decision-making

Messages · Evidence · Visuals · Speaker notes for each slide
04

Writing and editing

Reports, emails, documents, proposals, and proofreading

A ready-to-use document and key editorial decisions
05

Data analysis

Metric definitions, comparisons, root-cause analysis, and decision reports

Reproducible analysis, limitations, and decision recommendations
06

Research · comparison · purchasing

Research on products, technology, travel, services, and options

Sourced comparisons and conditional recommendations
07

Meetings · summaries

Organizing meeting notes, long documents, decisions, and action items

Evidence-based summaries, decisions, and action lists
08

Marketing and customer communication

Campaigns, product descriptions, support replies, and calls to action

Channel-appropriate messages and verifiable calls to action
09

Translation and localization

Preserving meaning and format in documents, UI, and marketing copy

Reviewable translation with terminology and uncertainty notes
10

Image creation

Design advertising, educational, and product images through scenes and formats

An image prompt ready to copy and use, with review criteria
11

Video production

Plan educational, product, and campaign videos using timelines and scene transitions

A video prompt ready to copy and use, with a shot review sheet

PREREQUISITE CHECK

Three things to check before reading

This is not a test of memorized answers. Think about each question first, then open the explanation to review the foundational concepts used in this course.

1Does writing a prompt actually carry out the task?

This tool only helps you write a request. Sending it to an external service and executing the task are separate actions, and you must check the required permissions separately.

2How should you record owners or deadlines that are missing from the source?

Do not guess to make the output look complete. Mark the information as unconfirmed, and ask a clarifying question when it is needed for the decision.

3If the table has the right shape, is the answer accurate too?

Format and meaning are separate things to check. Verify not only the required columns but also whether the facts, figures, and decision status from the source were preserved.

TEXTBOOK GUIDE

Main text that covers each concept from its background to the criteria for judging it

We explain the material section by section so readers new to IT can connect causes and effects without memorizing terms.

CONCEPT FLOW

How the chapters connect

The chapters are not isolated short answers to memorize. Follow them from left to right to see how each chapter's concepts support the next decision.

  1. 1.Turn a duplicate-save bug into a reproducible repair request
  2. 2.In quote comparisons, price conditions come before recommendations
  3. 3.Keep proposals distinct from confirmed decisions in meeting summaries
  4. 4.Align denominators and collection scope before comparing conversion rates
  5. 5.State preservation requirements before fluency preferences in translation requests
Prompt design for work: the overall map. If you lose track while reading the detailed explanations and chapters below, return to this sequence.

CONTROLLED EXPLANATION

Flow for finding an assumed approval in a meeting summary and rechecking

This author-created diagram follows a fictional note containing only a proposal: withhold the first output’s approval claim, revise the instruction, and recheck against the same source. It does not represent a measured model execution or a security guarantee.

Current state: 1 · Source input

Flow for finding an assumed approval in a meeting summary and rechecking

This author-created diagram follows a fictional note containing only a proposal: withhold the first output’s approval claim, revise the instruction, and recheck against the same source. It does not represent a measured model execution or a security guarantee.

Apply initial instructionsUnsupported approval foundFix the cause of failureKeep input conditions fixedAny mismatchBoth checks pass11 · Source input22 · Initial output33 · Held for evidence mismatch44 · Revise the instruction55 · Recheck the same source66 · Example judgment passed
  1. 1 · Source input

    September 20 launch proposed. Security review unfinished. No approval recorded.

  2. 2 · Initial output

    Faulty simulated output: September 20 launch approved.

  3. 3 · Held for evidence mismatch

    No approval evidence exists in the source. Do not use this as a confirmed plan.

  4. 4 · Revise the instruction

    Separate proposals from decisions, and mark approvals and deadlines that are not stated as unconfirmed.

  5. 5 · Recheck the same source

    Check both: no added approval and no inferred deadline.

  6. 6 · Example judgment passed

    Preserve only the proposal date; approval is unconfirmed. Evaluate new cases separately as well.

1 → 2
Apply initial instructions
2 → 3
Unsupported approval found
3 → 4
Fix the cause of failure
4 → 5
Keep input conditions fixed
5 → 3
Any mismatch
5 → 6
Both checks pass

The arrows show the review order. A failed recheck goes back to hold. Passing means meeting the criteria of this fictional example, not accuracy across an entire real service.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.

Key concepts that guide judgment

Conceptual explanation 01

A goal describes the work state that should change after the response

A request to create a report names an artifact but leaves its decision unspecified. A reviewer approving a purchase needs different evidence from an engineer narrowing an incident cause. First write one sentence stating what the reader should decide or do after reading the response.

In a fictional GPU expansion task, define the goal as comparing approval conditions for options A and B rather than introducing equipment. Budget, the current bottleneck, and each option’s constraints then take priority over a performance overview. A list of specifications that cannot support the reader’s decision does not meet this goal.

Separate the goal from permission to act. A request to prepare approval material does not authorize ordering equipment. This tool only builds the final prompt; sending the copied text to another service or carrying out the real task does not happen automatically.

When a task has several goals, check whether one deliverable can support them together. Incident recovery and quarterly investment planning require different evidence and approvers, so separate requests may be clearer. After defining the goal, identify the decision served by each output item and reconsider items with no such connection.

Why does this happen?
Defining the purpose helps identify which information could change the response.
When is it a problem?
If an overview is complete but provides no basis for approval, revisit the goal definition.
Common beginner misconceptions
It is a misconception that assigning an expert role also conveys the task's purpose. State the role and the desired decision separately.
How to verify it yourself
Underline the deliverable and the reader’s next action separately in your request.
To summarize this sectionWrite the reader’s decision and the limit of action authority alongside the deliverable name.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.
Conceptual explanation 02

Provide context as clearly separated facts that can change the conclusion

Context is the information needed for a decision, not the amount of pasted text. Separate current observations, scope, reference dates, and source material so it is clear what can be treated as established fact. The same number can refer to different periods or populations, so state each value with its conditions.

In a fictional conversion analysis, 80 buyers alone provide no denominator for a rate. Supplying 1,000 product-detail visitors for the same period and a deduplication rule allows an 8% calculation under that definition. An 8% figure using visitors from another period is not the same metric, even if it has the same name.

Put source quotations and your instructions in separate areas. For example, place meeting notes in a source area and ask for decisions and proposals to be distinguished in the instruction area. Headings and delimiters aid interpretation, but they are not themselves a security mechanism that neutralizes instructions inside the source.

Require clarification only when missing context could change the decision. Whether a price includes tax can change a budget decision, while a preferred report color may not matter to that decision. Instead of filling unconfirmed items with guesses, connect each necessary question to the judgment its answer could change.

Why does this happen?
The same statement means something different when its target or period changes.
When is it a problem?
If the response compares different periods or invents a denominator, inspect the input contract.
Common beginner misconceptions
More material does not automatically fill gaps. A long source remains incomplete if a critical condition is missing.
How to verify it yourself
Write the population, period, and unit beside each example value, and mark values you could not verify as unconfirmed.
To summarize this sectionKeep values, scope, dates, and sources together, and separate instructions from material.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.
Conceptual explanation 03

Success criteria turn plausibility into observable checks

A request to be accurate and professional is not enough to judge failure. Break success criteria into values to check in the answer, facts to preserve, and actions not to allow. Only by checking the impression of the wording separately from factual agreement can you filter out answers that look good but are wrong.

In the meeting-notes example, requiring an owner and deadline for every action item can actually encourage fabrication. A more accurate requirement is to use only the owners and deadlines that appear in the source and mark missing values as unconfirmed. Do not conclude that fidelity to the source improved just because there are fewer blanks.

Separate automated checks from human review. Preserved numbers and required columns are relatively easy to compare, whereas deciding whether a proposal became a decision requires reading the context. Passing one check does not mean the response passes the remaining criteria.

Write pass conditions and hold conditions in pairs. For example, when there is no budget evidence, allow only a conditional comparison and hold off on a final purchase recommendation. After receiving the answer, record pass, fail, or unconfirmed for each condition, and keep the part of the output that supports each verdict.

Why does this happen?
Defining observable conditions first lets you compare responses against the same criteria.
When is it a problem?
If the response invents owners or figures to fill required fields, the success criteria may be poorly designed.
Common beginner misconceptions
It is a misconception that complete formatting means the content has been verified. Check format and meaning separately.
How to verify it yourself
Use an example with no deadline in the source and check that the result leaves it unconfirmed.
To summarize this sectionPass an answer that matches the evidence, not one that is merely filled in.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.
Conceptual explanation 04

Split constraints into prohibitions, approvals, and stop conditions

Rather than the shape of the result, constraints define the boundaries the work must not cross. Distinguish material that must not be changed, actions allowed only after approval, and points where work must stop if evidence is missing. Putting all three into a single sentence that says to be careful makes it hard to know when a decision is actually needed.

For a fictional code fix, separate preserving the public API, prohibiting data deletion, and requiring approval before deployment. Fixing the bug does not authorize changing the API or deploying to production. The report should distinguish actions awaiting approval from changed files and completed checks.

Do not leave conflicting constraints for the model to resolve silently. Requiring every supporting detail while limiting the output to one sentence may remove necessary information. State which requirement takes priority and when the length limit should be reconsidered.

A prohibition in a prompt does not replace permission limits in the execution environment. Writing that delete commands are forbidden does not remove the tool's write access. In real automation, grant only the permissions needed and enforce approval and validation in the execution path; the browser activities in this course do not execute real tasks.

Why does this happen?
With only an outcome goal, actions outside the permitted scope may still be proposed as ways to achieve it.
When is it a problem?
If a revision request expands into ordering, sending, or deployment, recheck the approval boundary.
Common beginner misconceptions
Repeating a prohibition does not enforce access control. Tool permissions and approval procedures are needed separately.
How to verify it yourself
Split allowed work and work awaiting approval into two columns, and place each action in one of them.
To summarize this sectionCheck prohibited actions, actions requiring approval, and stopping points separately.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Safety best practices · 2026-09-14 · Adversarial-input testing and human review principles. The fictional records do not execute attacks or call models.
Conceptual explanation 05

An output contract also includes rules for how the reader interprets values

An instruction to answer in a table does not define what the columns mean. If monthly and annual costs are mixed in the same cost column, the table may look tidy but the comparison is wrong. Define column names, units, periods, and how to mark missing values together.

For a fictional quote comparison, use columns for option, monthly cost, inclusions, exclusions, and reference date. For an option whose cost is undisclosed, enter unconfirmed rather than zero. Zero asserts that there is no cost, whereas unconfirmed means there is no data to judge from; the two are not interchangeable.

Fixed field names alone are insufficient when requesting machine-readable JSON. Valid JavaScript Object Notation (JSON) syntax and the business validity of its values require different checks. The API’s Structured Outputs feature helps produce output matching a supported schema, but does not automatically correct factual mistakes.

This prompt tool does not enforce an API schema or validate result JSON. Users specify a format in text and must separately check its form and values in the service where they use it. If the task needs uncertainty explained, also provide space outside a compact table for evidence and reasons to withhold judgment.

Why does this happen?
Shared interpretation rules make results comparable and usable in subsequent work.
When is it a problem?
If an absent cost becomes zero or monthly and annual units are mixed, inspect the output contract.
Common beginner misconceptions
JSON formatting does not make facts true. A valid structure can still contain incorrect values.
How to verify it yourself
Enter an input with no price and check that the unconfirmed label and units are preserved.
To summarize this sectionDesign the format, units, unconfirmed-value notation, and meaning checks together.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Structured model outputs · 2026-09-14 · The distinction between schema conformance and content accuracy. This site does not execute the API’s Structured Outputs feature.
Conceptual explanation 06

Missing evidence should reduce confidence in the conclusion

A confidently worded model response is not the same as a verified fact. Even when a source link is present, read the document to check whether it supports the claim. Connecting each central claim to a source passage, applicable conditions, and verification status shows what needs review.

Suppose an official price list in a fictional purchase study gives a monthly fee but leaves tax and regional conditions unspecified. Those missing conditions need verification before the total cost can be finalized. Separating the subtotal supported by current material from unverified costs avoids inventing figures to finish the conclusion.

Separate source-backed facts, inferences drawn from them, and assumptions used for calculations. For example, more incidents alone do not establish that a new version caused them. Check alternative explanations such as deployment timing, traffic changes, and missing observations before deciding the scope of the conclusion.

For requests that need up-to-date information, state the reference date and the sources to check. However, telling the prompt to search does not give the service a search capability. If there is no tool or data access, require the answer to report freshness as unconfirmed and not to list link checks it did not actually perform as completed.

Why does this happen?
The link to sources and observations supports a judgment more than confident wording does.
When is it a problem?
Stating that you reviewed material you could not access is a failure of evidence reporting.
Common beginner misconceptions
It is a misconception that attaching a source link completes citation verification. Check whether the link actually supports the claim.
How to verify it yourself
Pick one claim and find its source sentence and effective date; if either is missing, mark it as needing verification.
To summarize this sectionSeparate facts, inferences, and assumptions, and withhold claims that remain unverified.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.
  • OpenAI · Safety best practices · 2026-09-14 · Adversarial-input testing and human review principles. The fictional records do not execute attacks or call models.
Conceptual explanation 07

Representative-case evaluation checks whether a changed instruction reduces real failures

A single good answer is not enough to say that a prompt is stable. The same input can produce different results, and some errors do not show up in easy examples. Prepare normal cases that represent the actual work together with boundary cases that are likely to fail.

For a meeting-notes prompt, distinguish notes with a confirmed decision, notes with only proposals, and notes missing an owner. Before looking at any answers, define the acceptable results and the prohibited inferences for each case. An answer that inserts a confirmed decision into every case fails on the boundary cases, however smoothly it reads.

Use the same inputs and evaluation criteria when comparing a revision with the original. Changing the prompt, model or service settings, and source scope together makes their effects difficult to separate. Record the version and failed criteria, but do not turn a few successful trials into a guarantee of accuracy across the whole task.

Looking only at format pass rates misses regressions that fail to preserve facts. If a revision reduces missing deadlines but invents owners who do not exist, do not adopt it. Classify the causes of failure as ambiguous instructions, insufficient material, or errors in the evaluation criteria, and after revising, recheck both the same cases and separate new cases.

Why does this happen?
You can judge whether a prompt has improved only by checking failure types, not by picking one example that looks good.
When is it a problem?
If only the edited example passes while other inputs gain invented facts, suspect a regression.
Common beginner misconceptions
It is a misconception that duplicating many evaluation cases increases diversity. Include different failure conditions.
How to verify it yourself
Create one normal case, one missing-information case, and one conflicting-source case, and apply the same pass criteria to each.
To summarize this sectionCompare before and after against the same criteria, and withhold adoption if new failures appear.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.
Conceptual explanation 08

Do not treat instructions inside source material as authority to act

A document under review may contain imperative text unrelated to the task. A request to summarize it and a sentence inside it asking for secrets to be sent do not carry the same authority. Treat the material as an object of analysis and prevent embedded action requests from redefining the original task.

Suppose a fictional customer note asks for an authentication token to be sent to an external address before review. The summary needs the customer’s request and supporting evidence, not authentication secrets. Do not execute that sentence; identify it as a suspicious request within the material and review it within the original summarization scope.

Input minimization is a control you can apply before writing instructions. Remove names, account secrets, and full customer records that are unnecessary for reproduction, replacing them with role names or short fictional examples. Asking the response not to expose a secret after pasting it does not undo the disclosure already made.

This site assembles inputs into a prompt inside the browser and does not call a model. When sending the copied prompt to an external AI service, separately check that service’s data handling terms and your authorization to use it. Reinforce real automation security with access limits, tool-call validation, and necessary human approval as well as instructions.

Why does this happen?
Executing requests found inside the source while processing it crosses the task’s trust boundary.
When is it a problem?
If summarization introduces external transmission or credential requests, inspect the boundary between material and instructions.
Common beginner misconceptions
Delimiter tags do not completely prevent prompt injection. They help interpretation and do not replace access control.
How to verify it yourself
Mark suspicious action requests in a fictional document and retain only information required for the original task.
To summarize this sectionSeparate source analysis from authorization to act, and remove unnecessary secrets before input.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Safety best practices · 2026-09-14 · Adversarial-input testing and human review principles. The fictional records do not execute attacks or call models.

CONCRETE CASES

Compare the failed and revised meeting-notes outputs against the source

The three rows use different fictional notes. Check whether each revision preserves the source’s certainty and missing information rather than filling every gap. The revised outputs are teaching references, not results from running a real model.

Compare the failed and revised meeting-notes outputs against the source
Fictional sourceFaulty outputReference revisionBasis for the verdict
The support document needs updating. No owner has been assigned.The support team lead owns the document update.Action: update the support document / Owner: unconfirmedThe existence of a task does not establish its owner. Do not add an owner absent from the source.
The product owner proposed including the feature in the next version. No agreement was reached.It was decided to add the feature to the next version.Proposal: include the feature in the next version / Decision: unconfirmedKnowing who made the proposal does not make it approved. Preserve the fact that no agreement was reached.
Security review is required. Its completion date has not been confirmed.Security review will finish by Friday.Required work: security review / Completion date: unconfirmedDo not turn a need into a deadline commitment. If there is no schedule, leave it as a question to confirm.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.
  • OpenAI · Safety best practices · 2026-09-14 · Adversarial-input testing and human review principles. The fictional records do not execute attacks or call models.

CHAPTER 1 / 5

Turn a duplicate-save bug into a reproducible repair request

In this fictional development case, clicking Save twice produces two network requests and the second response is 409. Supply the reproduction steps and responses without assuming that the frontend is the cause. The task is to inspect the duplicate-handling policy and propose the necessary change and regression checks.

Add the versions in use, relevant functions, and the expected normal save result. Require the public API to remain unchanged and prohibit database changes before approval. If the model cannot access the files, adjust the scope to proposed changes and places to inspect rather than completed implementation.

The reviewer checks a single click, two rapid clicks, and retry after the first request fails. Permanently disabling the button may prevent duplicates, but is not success if it prevents retry. Distinguish executed tests from tests that could not run, and retain a procedure for checking the same inputs after the change.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.
  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.

CHAPTER 2 / 5

In quote comparisons, price conditions come before recommendations

Fictional purchase material lists option A at 1.2 million KRW per month and option B at 12 million KRW per year. Option B converts to 1 million KRW per month, but that alone does not establish the better choice. First check contract length, tax, included support, and early termination terms.

In the prompt, state that only the provided prices may be used in calculations and that missing contract terms must remain unconfirmed. The comparison table should include not only cost but also whether mandatory requirements are met and which documents to check. Even if its price is lower, an option whose mandatory support requirements could not be confirmed is classified as needing further checking, not as recommended for approval.

The reviewer checks the conversion arithmetic separately from the recommendation. Correctly dividing 12 million KRW by 12 does not mean monthly payments are available. Distinguish the calculated monthly equivalent from actual payment terms and retain unresolved items for the approver.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.
  • OpenAI · Structured model outputs · 2026-09-14 · The distinction between schema conformance and content accuracy. This site does not execute the API’s Structured Outputs feature.

CHAPTER 3 / 5

Keep proposals distinct from confirmed decisions in meeting summaries

A fictional note says the product owner proposed launching next Friday and the security reviewer replied that more review was needed. A poor summary turns this into a confirmed launch next Friday. Because the source does not establish agreement, the correct classification is proposed or on hold, and the confirmed schedule is unconfirmed.

Structure the request to extract decisions, proposals, reasons for holds, and actions into separate columns. Fill owners and deadlines only when the source states them, linking a short supporting excerpt. Prohibit guessing missing roles or dates merely to make the table look complete.

Test both an ordinary detailed note and a note containing no decisions. Reporting no decisions for the latter is faithful to the source, not a task failure. Compare additions and omissions in decision status, owners, and deadlines before and after revision to ensure a polished summary has not changed the meaning.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Safety best practices · 2026-09-14 · Adversarial-input testing and human review principles. The fictional records do not execute attacks or call models.
  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.

CHAPTER 4 / 5

Align denominators and collection scope before comparing conversion rates

A fictional table shows 80 buyers and 1,000 visitors last week, then 72 buyers and 800 visitors this week. With the same definitions and period length, conversion rises from 8% to 9% while buyer count falls. Do not immediately interpret the higher rate as higher revenue or overall improvement.

Specify visitor deduplication, timezone, bot exclusion, and known collection gaps. If mobile visit events are missing this week, the smaller denominator may have inflated the rate. In that situation, withhold a causal conclusion and first request checks to establish the missing-data scope.

Set the output contract to separate the formula, observed change, possible explanations, and follow-up checks. A 1 percentage point increase and a 12.5% relative change are different expressions, so keep them distinct. Learners should review not only the arithmetic but also the comparison conditions and the evidence behind causal claims.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.
  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.

CHAPTER 5 / 5

State preservation requirements before fluency preferences in translation requests

A fictional installation document requires stopping the service before changing settings and specifies a 30-second wait. Even fluent translation changes the procedure and operational risk if it drops the stop condition or changes the wait to 30 minutes. Require comparison of numbers, units, negation, and prerequisites against the source.

Mark product names and commands as do-not-translate items and keep them separate from explanatory sentences. If a command in the source looks wrong, have the translator report the source problem and a suggested fix separately instead of quietly correcting it. Decide whether the user asked only for meaning preservation or also for technical correction, so responsibilities do not get mixed.

Use review columns for source, translation, preservation requirements, and open questions. Do not pass a translation that turns “must” into “need not,” even if the numbers match. After a correction, reread the surrounding conditions as well as the edited sentence, and do not mark a sentence as reviewed until its meaning has been checked.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.
  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.

Diagnose and re-review failed prompts

INTERACTIVE LAB 1 / 2

Lab 1 · Lab 1 · Repair a meeting summary that invents approval

Fictional source: the product owner proposed a September 20 launch. The security reviewer replied that review was unfinished. The faulty input asks, “Fill an owner and deadline for every item and summarize a confirmed plan.” The simulated faulty output says, “September 20 launch approved; security reviewer to finish the day before.” Neither approval nor a review completion date appears in the source.

Choose a revision that does not infer approval or deadlines. Compare it with the reference output in the explanation and identify the two checks to repeat against the same source. This activity evaluates supplied fictional records and does not call a real model.

Choose your verdict, then compare it against the reasoning.

Reference choice: B

A. Make the table more concise and keep the launch approval sentence.This changes only length, leaving the unsupported approval claim intact. The recheck fails because no source evidence supports approval.

B. Separate decisions from proposals, and mark missing owners and deadlines as unconfirmed.The reference output after revision is “Proposed launch: September 20 / Approval: unconfirmed / Hold reason: security review unfinished / Review completion date: unconfirmed.” It passes when a recheck against the same source confirms that no approval was added and no missing deadline was invented. This is a pass on an example judgment, not a measurement of real service output.

C. Add an expert project manager role and keep the other requirements the same.A new role leaves the faulty requirement to fill every gap with a definite value intact. Change the source-evidence and unconfirmed-value rules.

D. Delete all schedule language and omit approval status.This may reduce fabrication, but it also removes the proposal and the reason for the hold that the reader needs to know. It fails the goal of preserving the source.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.
  • OpenAI · Safety best practices · 2026-09-14 · Adversarial-input testing and human review principles. The fictional records do not execute attacks or call models.

INTERACTIVE LAB 2 / 2

Lab 2 · Lab 2 · Repair an approval decision for quotes with different units

Fictional material: option A costs 1.2 million KRW per month and B costs 12 million KRW per year; whether either includes tax is unconfirmed. The faulty input asks, “Recommend the option with the smaller number immediately.” The simulated faulty output says, “Approve A because 120 is smaller than 1,200.” The budget cap is a monthly equivalent of 1.1 million KRW and must be judged using the tax-inclusive total.

Choose a revision that handles both unit conversion and withholding approval. Identify what still cannot be concluded without tax conditions and compare your judgment with the recheck criteria in the explanation.

Choose your verdict, then compare it against the reasoning.

Reference choice: C

A. Keep A because its number is smaller and add a professional explanation.The error of comparing figures for different periods remains. Changing the tone does not repair the calculation basis.

B. Approve Option B as soon as it is calculated at 1 million KRW per month.The monthly conversion is correct, but the tax-inclusive total is unconfirmed. Final approval is premature because a condition required for the budget decision is missing.

C. Compare monthly equivalents but withhold approval until tax inclusion is verified.The reference output after revision is “A: 1.2 million KRW per month / B: monthly equivalent of 1 million KRW / Tax-inclusive total unconfirmed / Approval withheld.” Recheck separately the calculation, 1,200÷12=100 in units of ten thousand KRW, and the decision to withhold approval because of unconfirmed conditions. This does not mean payments are actually made monthly.

D. Assume there is no tax and omit the assumption label.This turns a condition that was not provided into an established fact. Even if the calculation looks polished, it fails because it goes beyond the supplied evidence.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.
  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.

KEY TERMS

Key terms in this unit

A goal describes the work state that should change after the response
Write the reader’s decision and the limit of action authority alongside the deliverable name.
Provide context as clearly separated facts that can change the conclusion
Keep values, scope, dates, and sources together, and separate instructions from material.
Success criteria turn plausibility into observable checks
Pass an answer that matches the evidence, not one that is merely filled in.
Split constraints into prohibitions, approvals, and stop conditions
Check prohibited actions, actions requiring approval, and stopping points separately.
An output contract also includes rules for how the reader interprets values
Design the format, units, unconfirmed-value notation, and meaning checks together.
Missing evidence should reduce confidence in the conclusion
Separate facts, inferences, and assumptions, and withhold claims that remain unverified.
Representative-case evaluation checks whether a changed instruction reduces real failures
Compare before and after against the same criteria, and withhold adoption if new failures appear.
Do not treat instructions inside source material as authority to act
Separate source analysis from authorization to act, and remove unnecessary secrets before input.

UNIT WORKBOOK

Exercises and worksheets for applying concepts to new situations

Start by checking basic principles, then expand to practical workplace decisions. After submitting an answer, you can see why every option is correct or incorrect, not just the correct answer.

Basic

Which instruction most clearly identifies the reader’s decision when requesting an incident report?

Choose your verdict, then compare it against the reasoning.

Reference choice: A

A. Compare observed evidence and remaining hypotheses so the owner can choose what to inspect next.This specifies the reader’s next action and the evidence needed for it. It does not demand an unsupported confirmed cause or completed recovery.

B. Answer like the world’s best expert.This specifies only a role and impression, not what the reader should decide.

C. Write a very long, exhaustive report.Length does not replace purpose, and the required coverage is unspecified.

D. Make every statement confidently definitive.This can turn unsupported hypotheses into facts. A confident tone is not verification evidence.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Prompt engineering · 2026-09-14 · Principles for separating instructions from source material and for evaluating prompts. The work examples and calculations were created by the author.

Apply

For equal-length periods, buyers/visitors changed from 80/1,000 to 72/800. Possible missing mobile visits remain unresolved in the later period. Which report is most appropriate?

Choose your verdict, then compare it against the reasoning.

Reference choice: C

A. Conclude definitively that revenue also increased because the conversion rate rose.There is no revenue data, and the visit collection conditions are unconfirmed. This infers both revenue and causation from a rate.

B. Report that conversion fell because buyer count decreased.Conversion rates must be calculated with their denominators. The arithmetic on the given figures is 8% to 9%.

C. Report the calculated rise from 8% to 9% but withhold an improvement conclusion until collection gaps are checked.This separates the arithmetic observation from data-quality judgment. A change of 1 percentage point is distinct from a verified explanation of its cause.

D. Ignore missing visitors because they do not affect the rate.Visitors form the denominator, so missing visits can change the rate. Checking collection conditions is part of the comparison.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.

Capstone

A revised prompt passes the format check on ordinary meeting notes but invents approval in a note containing only proposals. What is the most appropriate action?

Choose your verdict, then compare it against the reasoning.

Reference choice: B

A. Apply it to all work immediately because the normal case passed.This ignores a semantic error in a boundary case. Passing a format check does not guarantee source fidelity.

B. Revise the approval-inference rule and withhold adoption until the same cases and a new hold case are retested.This preserves the observed failure and connects revision to regression checking. Retesting must judge factual preservation and format separately.

C. Remove the failing note from the evaluation set to increase the pass rate.The metric improves while the real failure remains. An evaluation missing representative work cannot justify adoption.

D. Add the unsupported approval to the reference answer.Changing the reference answer to match an output error defeats the fidelity check. Set the reference answer from the supplied material.

These explanations and fictional workplace scenarios were created by the author based on official sources.

  • OpenAI · Evaluation best practices · 2026-09-14 · Task-specific evaluation, boundary cases, and human review principles. This unit does not teach a particular evaluation API.

PERSONAL WORKSHEET

A learning worksheet you adapt to your own environment

Your input remains only on the current browser screen and is not stored or transmitted externally. Use categories and pseudonyms instead of actual sensitive information.

OFFICIAL SOURCES

Verify against official sources

Official sources reviewed on September 14, 2026. The workplace scenarios and answers were created by the author for learning.

PROMPT ENGINEERING FOUNDATIONS

A good prompt is a work specification, not a magic spell

Clear outcomes, relevant context, success criteria, and boundaries make results more reproducible than adding many impressive-sounding adjectives. Even for complex work, avoid repeating instructions and retain only information that changes the result.

  1. 01

    Start with the outcome

    Instead of simply writing “analyze this,” specify the decision to support and the deliverable you need.

  2. 02

    Provide facts and context

    Provide internal terms the model cannot know, the current state, the reference date, and source text as clearly separated inputs.

  3. 03

    Write the pass line

    List items that confirm completion and the conditions for failure, rather than aiming for an answer that merely looks good.

  4. 04

    State the boundaries

    Define what must not change, what needs approval, and when to ask questions or stop if uncertain.

  5. 05

    Define the output format

    Specify the reader, the length, any table, code, or slide structure, and the content that must be preserved.

  6. 06

    Test with representative cases

    A prompt is input to a probabilistic system. Compare and refine results using real work examples.

BEFORE AND AFTER

Writing at length is not the same as writing accurately

An improved prompt is not a sentence padded with decorative words but a document that separates the evidence the model needs for its judgment from the completion conditions.

Source: composed by the author based on OpenAI's official prompt engineering guidance
Before improvement

One vague sentence

Make our company's GPU expansion report professional and impressive.
  • No indication of who will read it or what they will decide
  • No budget, figures, sources, or reference date
  • No slide count or success criteria
After improvement

A task specification that supports decisions

Goal Have the CFO and CTO approve one of the two expansion proposals.

Context Provide the budget, current bottleneck, outage cost, and quotes A/B.

Success criteria Compare the cost of inaction and selection criteria numerically.

Constraints Does not invent numbers absent from the data, and flags uncertainty.

Deliverable 10 slides, with a key message, diagram, and speaker notes for each slide.

PRE-SEND REVIEW

Check before sending

Even a completed prompt does not automatically produce a correct answer you can trust unchanged. Check the items below and evaluate actual results using representative examples.

  1. GoalIs the outcome the model must deliver expressed as actions and deliverables?
  2. EvidenceAre dates, figures, and internal context separated into facts and opinions?
  3. Pass lineIs there a condition the reader can use to confirm that the answer is complete?
  4. BoundariesAre security, permissions, prohibitions, and conditions for asking questions or stopping specified?
  5. EfficiencyHave you avoided repeating the same instructions and unnecessary examples?
  6. Personal dataHave secrets and personal information been replaced with pseudonyms or the minimum necessary information?

OFFICIAL OPENAI SOURCES

Verify against the official guidance

Reviewed August 27, 2026. Model behavior and recommended practices can change, so also check the latest content at the links.