Build a Good-Prompt Lab

Create and test a reusable planning prompt in ChatGPT.

Introduction

30 Second Summary

When you ask for help with a busy week, a vague request often returns a plan that ignores the deadline or the time you actually have. The result sounds helpful until you try to use it.

In this project, you will create a small lab in ChatGPT for testing prompts. You will use a fixed rubric to turn a vague planning request into a reusable prompt for work or study.

What You'll Build

Your finished Prompt Lab will place a weak weekly plan beside an improved one, with a scorecard showing exactly why the stronger prompt is more reliable.

By the end of this project, you'll have:

A visible bad-prompt response you can score against four success criteria to expose missing priorities, deadlines, time limits, and format requirements.
A simple inference map you can use to explain how prompt text becomes tokens, conversation context, next-token prediction, and a visible response.
A reusable Outcome-Context-Constraints-Format prompt that you can test against changing deadlines and time limits with a three-case scorecard.
Secret Mission: Add one compact few-shot example that fixes the weakest edge case without making the final prompt unnecessarily long.

Are there any prerequisites?

You only need a Mac with a web browser. ChatGPT Free costs $0, though usage limits can apply.

Before We Start

A useful prompt starts with a real reason to improve the response. This is your moment to choose a genuine but non-sensitive work or study planning task before the hands-on experiments begin.

Open Your Prompt Lab

Prompt comparisons only mean something when every test starts from a clean workspace. A saved conversation keeps your experiments together.

This step sets up ChatGPT as your prompt-testing lab. A fresh conversation prevents unrelated chat history from shaping your first result.

In this step, get ready to:

Sign in to ChatGPT with free account access.
Create a fresh saved conversation for your Prompt Lab.
Prove the Prompt Lab follows an exact instruction.

Open ChatGPT in your browser

A signed-in account keeps your Prompt Lab available in chat history. This gives every later experiment one saved home.

Account setup can feel like a detour. The free plan is enough for every experiment in this project.

Navigate to https://chatgpt.com/ in your web browser.
Follow ChatGPT's account prompts to sign in if you already have an account.
Follow ChatGPT's account prompts to create a free account if you do not have one.
Confirm ChatGPT no longer asks you to authenticate.

That is the account hurdle cleared. Your experiments can now stay available in chat history.

Cannot access your account?

Return to the ChatGPT home page before retrying your account method.
Check your inbox if ChatGPT asks you to verify your account.

Still stuck? Help me access my ChatGPT account for a prompt-testing project.

Start a fresh Prompt Lab conversation

Conversation context includes the messages already present in a chat. A fresh conversation removes unrelated messages from your comparison.

Keep the Lab Non-Sensitive

ChatGPT chats may be reviewed or used to improve AI models. Your prompt experiments do not need confidential details.

Use fictional or non-sensitive work and study details throughout the Prompt Lab.

Use the conversation control at the top of ChatGPT's left sidebar to start a fresh conversation.
Designate this conversation as your Prompt Lab for the rest of the project.
Confirm the message composer is empty.
Confirm no earlier messages appear above the composer.

You now have a blank conversation with no unrelated messages. That clean baseline makes later comparisons meaningful.

Send a readiness check

Before you send the check, what exact text do you expect ChatGPT to return?

Paste this readiness prompt into the message composer:

Reply with exactly: Prompt Lab ready.

What Does This Prompt Test?

The phrase Reply with exactly: gives ChatGPT a precise output constraint.
The text Prompt Lab ready. provides a literal result that you can check.

Send the readiness prompt using the arrow control beside the message composer.

You should see a response containing Prompt Lab ready. Your Prompt Lab can now accept an instruction and return a checkable result.

Confirm the conversation appears in ChatGPT's chat history sidebar.

Your saved conversation now contains the readiness prompt plus its response. That gives you a consistent workspace for every comparison ahead.

Response Contains Extra Words?

Check that your prompt matches Reply with exactly: Prompt Lab ready. character for character.
Start another fresh conversation if the current chat contains any earlier instructions.
Send the readiness prompt again in the fresh conversation.

Need another pair of eyes? Help me diagnose why ChatGPT added extra text to my readiness response.

Your Prompt Lab is saved with a verified starting point. Next, you will test a deliberately vague request and measure where it falls short.

Make a Bad Prompt Fail

Your Prompt Lab is ready. Now you will create a baseline response that exposes what a vague prompt leaves to chance.

ChatGPT only sees one broad request in this test. The request omits the priorities, deadlines, available time, and output format behind the target plan.

A fixed evaluation rubric gives those gaps a measurable score. This turns an unreliable feeling into visible evidence.

In this step, get ready to:

Establish a four-point test for a fictional planning target.
Run the vague baseline prompt in a separate fresh chat.
Capture a scored diagnosis with at least two shortcomings.

Define the hidden test

The hidden target lives in Prompt Lab. The baseline chat receives none of it.

Add the fictional scenario and fixed rubric to the Prompt Lab conversation from earlier by sending the note below:

Hidden test scenario

Finish a 1,200-word report by Friday, prepare for a Thursday quiz, and work within 90 minutes per weekday from Monday through Thursday.

Success rubric

1. Pass/Fail: Prioritize the Thursday quiz as the earlier deadline.
2. Pass/Fail: Stay within 90 minutes per weekday.
3. Pass/Fail: Allocate time to both the report and quiz.
4. Pass/Fail: Present a day-by-day plan.

Why define the rubric first?

A fixed rubric keeps the test focused on observable evidence. Storing it in Prompt Lab keeps the baseline chat free of clues.

Your experiment now has a fair target. The hidden scenario and all four criteria are visible in your saved Prompt Lab conversation.

Run the vague baseline

Fresh chats isolate conversation context. The baseline chat starts with no clues from your earlier work.

Use the new-conversation control in ChatGPT's left sidebar to start a separate fresh chat.

Before you send it, which details do you think ChatGPT will invent or leave vague?

Submit the vague request by sending only this prompt:

Help me plan my week.

Why does this response fall short?

The request gives ChatGPT one broad objective. It supplies no task names or limits.

The response must choose assumptions or ask for more detail. That shortfall is the designed result.

You'll see a generic weekly plan or a request for more detail. Either result lacks the hidden scenario's task-specific requirements.

Leave the response unchanged in this fresh chat.

Does the chat contain earlier context?

Start another fresh chat if messages appear above the baseline request.
Send only the baseline request in the replacement chat.

Still unsure whether the test is isolated? Help me check my baseline chat.

Score the response

The rubric converts a variable response into a repeatable score. A criterion passes only when the response visibly meets it.

Return to the saved Prompt Lab conversation from earlier.

Before you score it, how many criteria do you think the vague request satisfied by chance?

Compare the baseline response with the four criteria saved in Prompt Lab.
Mark each criterion as Pass or Fail.
Identify the two most important shortcomings.
Complete this evidence scorecard in Prompt Lab:

Baseline results

1. Prioritize the Thursday quiz as the earlier deadline: [[THURSDAY_QUIZ_RESULT="Pass or Fail"]]
2. Stay within 90 minutes per weekday: [[TIME_LIMIT_RESULT="Pass or Fail"]]
3. Allocate time to both the report and quiz: [[BOTH_TASKS_RESULT="Pass or Fail"]]
4. Present a day-by-day plan: [[DAY_PLAN_RESULT="Pass or Fail"]]

Baseline score: [[BASELINE_SCORE="0 to 4"]]/4

Two most important shortcomings

1. [[SHORTCOMING_ONE="describe the first shortcoming"]]
2. [[SHORTCOMING_TWO="describe the second shortcoming"]]

What counts as a pass?

Award Pass only when the response visibly satisfies the full criterion. Use Fail for missing, vague, or conflicting evidence.

Send the completed scorecard in Prompt Lab.

You should see at least two Fail results. Those failures show which missing details weakened the baseline.

Seeing fewer than two failures?

Confirm that the baseline chat contains only the original baseline request.
Score missing evidence as Fail even when the response sounds polished.

Need help applying the rubric? Help me score the baseline consistently.

That experiment paid off: you have a scored baseline with at least two visible gaps. Next, you'll trace the missing information behind the baseline's assumptions.

Map the AI's Inferences

Your vague weekly prompt in ChatGPT exposed at least two gaps in the baseline plan. Those misses came from details the prompt left for the model to infer.

The model processes prompt text as tokens. Conversation history supplies context.

Learned patterns influence likely next-token choices. Generated tokens form the visible response.

In this step, get ready to:

Build a simplified map of how the AI produces a response.
Run a diagnostic prompt that reveals three plausible interpretations.
Classify the missing details as Outcome, Context, Constraints, or Format.

Build the simplified inference map

An inference map gives each stage a clear place. It describes observable inputs plus outputs without presenting a transcript of hidden reasoning.

Switch back to the Prompt Lab conversation from earlier.
Save the simplified inference map in that conversation by sending this text:

Simplified inference map:
1. Prompt text is split into tokens.
2. Conversation history provides context.
3. Learned patterns influence likely next-token choices.
4. Generated tokens form the visible response.

How to Read the Map

Tokens are chunks that represent commonly occurring character sequences.
Context is the conversation history available to the model.
Learned patterns influence which token is likely to come next.
Generated tokens combine to form the response you can read.

Confirm the four-stage map is visible as your latest message in the Prompt Lab.

Diagnose the vague prompt

The original request can support several sensible meanings. The diagnostic prompt makes that ambiguity visible.

Before you send the diagnostic prompt, what different kinds of weekly plans could the vague request produce?

Send the diagnostic prompt in the same Prompt Lab conversation:

Explain why “Help me plan my week” is ambiguous using this simplified model: text is split into tokens; the model uses conversation context and learned patterns to predict response tokens one at a time. List three plausible interpretations. Do not claim to reveal hidden reasoning.

What Does This Prompt Test?

The opening request focuses the response on ambiguity.
The simplified model limits the explanation to the concepts in your inference map.
The three interpretations make the prompt's open choices visible.
The final instruction prevents claims about hidden reasoning.

Read the response to confirm it includes the terms tokens, context, and next token.
Count the plausible interpretations to confirm the response gives three.

That is the key diagnosis: the vague prompt leaves multiple reasonable targets open.

Missing a Term or Interpretation?

Check that you sent the diagnostic prompt exactly as shown.
Resend the exact prompt if the response provides fewer than three interpretations.
Keep extra instructions out of the message because they can shift the response away from the diagnostic task.

Still missing part of the diagnostic response? Help me check why my diagnostic prompt did not produce the required terms and three interpretations.

Classify the missing details

The four labels turn each vague area into a specific prompt-design problem. Each label captures a different kind of information the baseline needed.

What Each Label Captures

These labels are a teaching scaffold for this prompt lab.

Outcome captures the result the response should help you achieve.
Context captures background facts that can change the plan.
Constraints capture boundaries that the plan must respect.
Format captures the visible structure of the answer.

Write Outcome: followed by the two results the weekly plan must achieve.
Write Context: followed by the report length plus both deadlines.
Write Constraints: followed by the 90-minute weekday limit from Monday through Thursday.
Write Format: followed by the requested day-by-day plan structure.
Match each plausible interpretation to the missing label that allowed it.
Write one comparison sentence for each interpretation in the Prompt Lab.

Before you check the complete diagnosis, which label do you expect to explain most of the baseline failures?

Scan the diagnostic response for the exact terms tokens, context, and next token.
Count three plausible interpretations of the vague weekly planning request.
Confirm that Outcome, Context, Constraints, and Format each contain a missing scenario detail.
Check that every interpretation has a comparison sentence tied to one of the four labels.

Strong work. Your Prompt Lab now shows how missing information widens the set of plausible responses.

Your diagnosis now pinpoints the information that the vague prompt left open. Next, you'll turn those gaps into a structured planning prompt with a measurable result.

Build and Compare a Good Prompt

Your baseline exposed the assumptions ChatGPT had to make. Your inference map showed where the vague request left gaps.

A structured prompt replaces guesswork with a clear outcome, relevant context, meaningful constraints, and an observable output format. Those four sections give ChatGPT a much narrower target.

In this step, get ready to:

Build an Outcome-Context-Constraints-Format prompt for the fictional planning scenario.
Test the improved prompt in a fresh conversation.
Save a compact bad-versus-good comparison in your Prompt Lab.

Build the structured prompt

Outcome-Context-Constraints-Format is a teaching scaffold for this lab. Each label gives one type of decision-changing detail a fixed home.

Copy the complete structured prompt below:

Outcome:
Create a realistic Monday-to-Thursday plan that helps me finish a 1,200-word report by Friday and prepare for a Thursday quiz.

Context:
I have 90 minutes available each weekday from Monday through Thursday. The quiz has the earlier deadline.

Constraints:
- Prioritize the Thursday quiz.
- Include time for both the report and quiz.
- Cover Monday through Thursday.
- Keep each day's total at or below 90 minutes.
- Ask at most one clarifying question only if missing information prevents a workable plan.

Output Format:
- Use a table with columns: day, task, minutes, completion check.
- Follow the table with exactly two short risk notes.

Why These Sections Work

Outcome names the result that the plan must support.
Context supplies the available time and the order of the deadlines.
Constraints protect the schedule from ignoring either task or exceeding the daily limit.
Output Format turns the response into evidence that you can check against the rubric.

Run the improved test

A fresh conversation removes the baseline response from the model's current conversation history. This keeps the comparison focused on the difference between the two prompts.

Use the new-chat control in the ChatGPT sidebar to start a fresh conversation.
Paste the structured prompt into the message box.

Before you send it, predict whether every rubric item will become easier to score.

Send the structured prompt.

You should see a Monday-to-Thursday table with day, task, minutes, and completion check columns. Every day's total should stay at or below 90 minutes.

The plan should include both tasks. It should prioritize the Thursday quiz.

Exactly two short risk notes should follow the table.

Add the minutes within each day to check the 90-minute limit.
Check that the plan includes time for the report and the quiz.
Confirm that the plan prioritizes the Thursday quiz.
Verify that the response uses the requested table columns and two risk notes.

That is the improved test working. Every requirement now has visible evidence in the response.

Does a Rubric Item Still Fail?

Check that all four labels and every constraint are present in the message you sent.
Start another fresh conversation.
Send the complete structured prompt again.

Still missing a requirement? Help me inspect my structured planning prompt.

Compare the results

Your improved test belongs beside the original evidence. Saving both results in the Prompt Lab keeps the experiment easy to revisit.

Return to the Prompt Lab conversation from earlier.
Save the structured prompt in your main record by sending:

Save the following as the improved prompt in this Prompt Lab. Do not answer the planning task.

Outcome:
Create a realistic Monday-to-Thursday plan that helps me finish a 1,200-word report by Friday and prepare for a Thursday quiz.

Context:
I have 90 minutes available each weekday from Monday through Thursday. The quiz has the earlier deadline.

Constraints:
- Prioritize the Thursday quiz.
- Include time for both the report and quiz.
- Cover Monday through Thursday.
- Keep each day's total at or below 90 minutes.
- Ask at most one clarifying question only if missing information prevents a workable plan.

Output Format:
- Use a table with columns: day, task, minutes, completion check.
- Follow the table with exactly two short risk notes.

What This Saves

This message places the complete improved prompt beside your earlier rubric and diagnosis. Your Prompt Lab now holds the prompt that produced the stronger test result.

ChatGPT should acknowledge the saved prompt without replacing its four labeled sections.

A fixed rubric makes the final comparison fair. Both prompts face the same four criteria.

Before you send the next message, predict how many Pass values the good-prompt column should contain.

Create the compact comparison from your saved evidence by sending:

Create a compact bad-versus-good comparison table for the four criteria already in this conversation.

Use the bad-prompt pass/fail scores already recorded above.
Record the good prompt as Pass for each criterion.
Use columns: criterion, bad prompt, good prompt.
Below the table, state the total pass count for each prompt.
Do not add new criteria.

What This Comparison Measures

The bad-prompt column preserves the evidence from your vague baseline.
The good-prompt column records the four checks you completed in the fresh chat.
The totals make the improvement visible without relying on exact response wording.

You should see four criteria rows. The good-prompt column should show Pass in all four rows.

The bad-prompt total should preserve at least two failures from your baseline. The good-prompt total should show four passes.

That is the comparison complete. Your Prompt Lab now shows exactly where the structured prompt beats the vague baseline.

Is the Comparison Table Incorrect?

Check that the original four-point rubric still appears earlier in the Prompt Lab.
Resend the comparison prompt if ChatGPT adds new criteria.

Need help matching the table to your evidence? Help me correct my prompt comparison.

Your structured prompt now has a four-pass result and a side-by-side comparison. Next, you will test whether it stays reliable when the planning conditions change.

Test, Score, and Refine

Your structured prompt already produced a plan that meets your first rubric. A single strong response can still be a lucky fit for one scenario.

You will now convert that prompt into a reusable template. Testing it in fresh ChatGPT conversations reveals how reliably it handles tighter time limits or missing information.

In this step, get ready to:

Convert the structured prompt into a reusable planning template.
Build a three-case scorecard from fresh test chats.
Save a targeted revision as Final Prompt.

Turn the prompt into a reusable template

A reusable template keeps the planning method stable while the scenario details become placeholders. This lets you test new tasks without rebuilding the prompt from scratch.

Switch back to the Prompt Lab conversation from earlier.
Save a reusable version of the structured prompt by sending this prompt:

Store this as Reusable Prompt Template. Do not generate a plan.

Outcome:
Create a realistic plan that completes [[TASKS="tasks to complete"]] by their deadlines.

Context:
[[RELEVANT_CONTEXT="relevant context that could change the plan"]]

Constraints:
- Deadlines: [[DEADLINES="deadline for each task, or deadline unknown"]]
- Available time: [[AVAILABLE_TIME="available minutes for each planning day"]]
- Prioritize the earliest stated deadline.
- Allocate time to every task.
- Never exceed the available time for any day.
- Ask exactly one clarifying question only if missing information prevents a workable plan. Otherwise, state one brief assumption and continue.

Output Format:
- Use a table with columns: day, task, minutes, completion check.
- Make every completion check observable.
- Follow the table with exactly two short risk notes.

How does this template stay reusable?

The four placeholders hold the tasks, deadlines, available time, and relevant context that change between planning scenarios.
The labeled sections preserve the planning goal and the rules that made your improved response successful.
The clarification rule gives ChatGPT one controlled way to handle information that blocks a workable plan.

Confirm that your sent message contains the four labeled sections.
Confirm that the template includes named placeholders for tasks, deadlines, available time, and relevant context.

Your Prompt Lab now contains a reusable template instead of a prompt tied to one week. The stable instructions give every test the same starting point.

Did ChatGPT create a plan immediately?

Your sent message still preserves the reusable template. Check that its first line tells ChatGPT to store the template without generating a plan.

Need help checking the template? Help me verify that my planning prompt has reusable placeholders and a controlled clarification rule.

Test three planning cases

Each test changes one important condition while preserving the same prompt structure. Fresh conversations keep earlier responses from influencing the comparison.

What are the three tests?

The original case tests whether the reusable version preserves your first success.
The tight-time case tests whether the daily limit still controls the plan.
The missing-deadline case tests whether the clarification rule handles incomplete context.

Start a fresh ChatGPT conversation using the new-chat control.
Run the original case by sending this prompt:

Outcome:
Create a realistic plan to finish a 1,200-word report by Friday and prepare for a Thursday quiz.

Context:
I can work from Monday through Thursday. The report must be 1,200 words.

Constraints:
- Use no more than 90 minutes per weekday.
- Prioritize the earlier deadline.
- Allocate time to both tasks.
- Ask exactly one clarifying question only if missing information prevents a workable plan.

Output Format:
- Use a table with columns: day, task, minutes, completion check.
- Make every completion check observable.
- Follow the table with exactly two short risk notes.

Why repeat the original case?

This test checks whether converting the prompt into a template changed its performance. It gives the two edge cases a stable comparison point.

Read the response against the same requirements you used for the improved prompt.
Keep this chat available as the original-case test.

You should see both tasks in a day-by-day table. Every weekday should stay at or below 90 minutes.

Does the original case miss a requirement?

Keep the response unchanged because a miss is useful scorecard evidence. Check that you sent the complete prompt in a fresh conversation.

Need help locating the miss? Help me compare this planning response with the original success requirements.

Start another fresh ChatGPT conversation using the new-chat control.
Run the tight-time case by sending this prompt:

Outcome:
Create a realistic plan to finish a 1,200-word report by Friday and prepare for a Thursday quiz.

Context:
I can work from Monday through Thursday. The report must be 1,200 words.

Constraints:
- Use no more than 45 minutes per weekday.
- Prioritize the earlier deadline.
- Allocate time to both tasks.
- Ask exactly one clarifying question only if missing information prevents a workable plan.

Output Format:
- Use a table with columns: day, task, minutes, completion check.
- Make every completion check observable.
- Follow the table with exactly two short risk notes.

What does the tighter limit test?

Only the available time changes from 90 minutes to 45 minutes per weekday. This isolates whether the constraint actually controls the schedule.

Check every day in the table against the 45-minute limit.
Keep this chat available as the tight-time test.

A compliant response keeps every day at or below 45 minutes. It still allocates time to the report and quiz.

Does one day exceed 45 minutes?

Keep the response as evidence of weak Constraint Compliance. Check whether the daily total is split across multiple rows before deciding that the limit passed.

Need help checking the totals? Help me audit each day's minutes against the 45-minute limit.

Before you run the missing-deadline test, do you think the prompt will ask one question or proceed with a brief assumption?

Start another fresh ChatGPT conversation using the new-chat control.
Run the missing-deadline case by sending this prompt:

Outcome:
Create a realistic plan to finish a 1,200-word report and prepare for a Thursday quiz.

Context:
I can work from Monday through Thursday. The report must be 1,200 words.

Constraints:
- The quiz is due Thursday.
- The report deadline is unknown.
- Use no more than 90 minutes per weekday.
- Prioritize the earliest stated deadline.
- Allocate time to both tasks.
- Ask exactly one clarifying question only if missing information prevents a workable plan. Otherwise, state one brief assumption and continue.

Output Format:
- Use a table with columns: day, task, minutes, completion check.
- Make every completion check observable.
- Follow the table with exactly two short risk notes.

What does the missing deadline reveal?

The report deadline is the only missing detail. The response reveals whether the clarification rule gives ChatGPT enough direction to handle that gap consistently.

Check whether the response asks no more than one clarifying question.
Check a completed plan for one brief assumption about the missing deadline.
Keep this chat available as the missing-deadline test.

The response should ask no more than one question. A response that proceeds should state one brief assumption without inventing a firm deadline.

Did the response invent a deadline?

Keep the response unchanged because the invented deadline identifies a Context Use weakness. Confirm that the prompt explicitly says the report deadline is unknown.

Need help diagnosing the response? Help me check how this plan handled the missing report deadline.

How does the scoring scale work?

Your evaluation rubric uses a three-point scale for each criterion. The four scores add up to a maximum of 8 for each case.

A score of 0 means the response misses the criterion.
A score of 1 means the response partially meets the criterion.
A score of 2 means the response fully meets the criterion.
Objective Fit measures whether the plan addresses every task with sensible priority.
Context Use measures whether the plan uses relevant details without inventing missing facts.
Constraint Compliance measures whether the plan respects every stated time limit and deadline.
Format Compliance measures whether the response follows the requested table and risk-note structure.

Assign four scores to each test response using the scale above.
Return to the Prompt Lab conversation from earlier.
Build the compact scorecard by sending this prompt:

Create a compact Three-Case Scorecard from my scores.
Use columns for Case, Objective Fit, Context Use, Constraint Compliance, Format Compliance, and Total out of 8.

Original case: Objective Fit [[ORIGINAL_OBJECTIVE_FIT="0, 1, or 2"]], Context Use [[ORIGINAL_CONTEXT_USE="0, 1, or 2"]], Constraint Compliance [[ORIGINAL_CONSTRAINT_COMPLIANCE="0, 1, or 2"]], Format Compliance [[ORIGINAL_FORMAT_COMPLIANCE="0, 1, or 2"]].
45-minutes-per-day case: Objective Fit [[TIGHT_OBJECTIVE_FIT="0, 1, or 2"]], Context Use [[TIGHT_CONTEXT_USE="0, 1, or 2"]], Constraint Compliance [[TIGHT_CONSTRAINT_COMPLIANCE="0, 1, or 2"]], Format Compliance [[TIGHT_FORMAT_COMPLIANCE="0, 1, or 2"]].
Missing-deadline case: Objective Fit [[MISSING_OBJECTIVE_FIT="0, 1, or 2"]], Context Use [[MISSING_CONTEXT_USE="0, 1, or 2"]], Constraint Compliance [[MISSING_CONSTRAINT_COMPLIANCE="0, 1, or 2"]], Format Compliance [[MISSING_FORMAT_COMPLIANCE="0, 1, or 2"]].

Keep my scores exactly as entered.
Calculate each total and the average of the three totals out of 8.
Flag any stated time-limit or deadline violation.

What does the scorecard prove?

The fixed criteria make three different responses comparable. The average shows whether the template performs reliably across the full test set.

Confirm that each case has a total out of 8.
Confirm that the scorecard shows the average of all three totals.
Check the scorecard for any flagged time-limit or deadline violation.

You now have visible evidence of the template's strongest case and weakest case. That evidence tells you where one targeted revision can help.

Do the totals look wrong?

Add the four criterion scores for each row yourself. Divide the sum of the three row totals by three to check the displayed average.

Need help checking the arithmetic? Help me verify the totals and average in my three-case scorecard.

Refine the weakest instruction

A targeted revision changes the instruction linked to the weakest score. Keeping every other instruction stable makes the effect of that change measurable.

Which instruction should change?

Low Objective Fit points to the Outcome instruction.
Low Context Use points to the relevant Context line or the missing-information rule.
Low Constraint Compliance points to the specific time-limit or deadline instruction.
Low Format Compliance points to the relevant Output Format instruction.

Find the case with the lowest total in your Three-Case Scorecard.
Use its lowest criterion to choose the responsible instruction.
Choose the response that required the most interpretation if two cases are tied.
Edit only the responsible instruction in your reusable template.
Place Final Prompt on the first line of the revised template.
Send the complete revised template in the Prompt Lab conversation.

The sent message now preserves your revised template under the name Final Prompt. Every unchanged line remains a control for the rerun.

Did more than one instruction change?

Compare the reusable template with Final Prompt line by line. Restore every unrelated change before running the next test.

Need help isolating the edit? Help me identify whether these prompt versions differ in more than one instruction.

Start a fresh ChatGPT conversation using the new-chat control.
Copy Final Prompt from the Prompt Lab conversation.
Replace its placeholders with the values from the lowest-scoring case.

Before you send it, do you expect the targeted instruction to improve the weak criterion without lowering the other scores?

Send the completed Final Prompt.

You should receive a new response for the same edge case. This gives you a direct before-and-after comparison for the single revised instruction.

Compare the rerun with the earlier response for that case.
Assign new scores for all four criteria.
Return to the Prompt Lab conversation from earlier.
Update the scorecard by sending this prompt:

Update the Three-Case Scorecard using my rerun scores.

Case rerun: [[RERUN_CASE="name of the rerun case"]].
Objective Fit: [[RERUN_OBJECTIVE_FIT="0, 1, or 2"]].
Context Use: [[RERUN_CONTEXT_USE="0, 1, or 2"]].
Constraint Compliance: [[RERUN_CONSTRAINT_COMPLIANCE="0, 1, or 2"]].
Format Compliance: [[RERUN_FORMAT_COMPLIANCE="0, 1, or 2"]].

Replace only the previous scores for this case.
Recalculate its total and the three-case average out of 8.
Flag any stated time-limit or deadline violation.

What completes the refinement loop?

Replacing only the rerun row preserves a fair comparison with the other two cases. The recalculated average measures the final template across all three conditions.

Confirm that the updated average is at least 6 out of 8.
Confirm that no case contains a stated time-limit violation.
Confirm that no case contains a stated deadline violation.

Is the average still below 6?

Rewrite the same targeted instruction with a more objective requirement.
Keep every other instruction unchanged.
Rerun the same weakest case in another fresh chat.
Update only that case's scores in the scorecard.

Need help strengthening one line? Help me make this instruction more testable without changing the rest of my prompt.

That is a strong finish. Your Final Prompt now holds up across three planning conditions without breaking a stated limit.

Secret mission

Teach the Prompt with One Example

Your Final Prompt already handles three planning cases. Now you will add one compact example that teaches its weakest behavior, then keep that example only if it protects the quality of both verification results.

Clean Up Your Resources

Everything you created is stored in ChatGPT conversations. ChatGPT Free costs $0, although usage limits can still apply.

Choose whether to keep your conversations available, pause by closing the browser, or delete them entirely.

Resources you used:

Your Prompt Lab conversation containing the rubric, inference map, reusable template, scorecard, Final Prompt, and verification notes.
Your Baseline Test Chat containing the deliberately vague prompt and its response.
Your Improved Prompt Test Chat containing the structured prompt and its response.
Your Test Case Chats covering the original scenario, reduced available time, and a missing deadline.
Your Few-Shot Verification Chats covering the original scenario and weakest edge case.

Keep everything running

Leave the conversations in your chat history if you want to reuse the Final Prompt or continue testing it.

Keep the Prompt Lab conversation in your ChatGPT history.
Keep the separate test chats for future comparisons.
Return to the Final Prompt whenever you need help planning a new work or study task.

Pause - I'll come back to this later

Closing the browser pauses your work while leaving every conversation available in your account.

Confirm that the Prompt Lab conversation appears in your ChatGPT history.
Close the browser window.
Use the same ChatGPT account when you return.

Delete - I don't want to use this again

Deleting the conversations removes your Prompt Lab and its supporting test results from your chat history.

Locate the conversation containing your Final Prompt in your ChatGPT history.
Open the conversation controls beside it.
Choose the option to delete the conversation.
Confirm the deletion.
Repeat the deletion process for the Baseline Test Chat.
Repeat the deletion process for the Improved Prompt Test Chat.
Repeat the deletion process for each Test Case Chat.
Repeat the deletion process for each Few-Shot Verification Chat.

You should no longer see the Prompt Lab or its test chats in your chat history.

Keep Future Tests Non-Sensitive

Fictional or non-sensitive scenarios keep confidential work, school, or personal information out of future prompt tests.

Nice Work!

You did it! Your ChatGPT Prompt Lab now contains a tested Final Prompt backed by visible evidence from several scenarios.

You've learned how to:

Build a bad-prompt baseline. Score its response with a four-point rubric to expose missing priorities, deadlines, time limits, and output requirements.
Map ambiguity through tokens, conversation context, and next-token generation. Rebuild the request with labeled Outcome, Context, Constraints, and Output Format sections.
Test a reusable template across three changing scenarios. Measure each response with a 0-to-2 scorecard. Refine the weakest case while preserving every stated deadline and time limit.
Secret Mission: Add one compact few-shot example to address a recurring weakness. Verify that the original scenario and weakest edge case follow its structure without copying its facts.

Ready to quiz yourself?