Build a ChatGPT Action Brief Assistant

Use ChatGPT Projects to turn messy notes into source-grounded action briefs.

Introduction

30 Second Summary

Messy notes can hide the decision that matters most. A polished summary can still leave you wondering who owns the next task or when it is due.

In this project, you will build an Action Brief Assistant inside ChatGPT Projects. It will turn scattered meeting, study, or planning notes into a structured brief without guessing missing details.

What You'll Build

Paste a page of scattered notes into your finished assistant and watch it return a concise brief that shows what was decided, what remains uncertain, and what needs to happen next.

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

A reusable six-section action brief that organizes every note set into a summary, confirmed decisions, tentative ideas, action items, open questions, and missing information.
Clear evidence boundaries that keep tentative suggestions separate from confirmed commitments. Missing owners or dates appear as Not stated.
A repeatable four-check scorecard that tests structure, grounding, certainty, and missing fields across different note sets.
Secret Mission: Create a trusted brief from your own low-sensitivity notes. Audit every claim before saving the result as a reusable project source.

Are there any prerequisites?

You need a Mac and a ChatGPT account. Safari is already included with macOS, so you do not need to install any software.

Before We Start

Before the hands-on work begins, this checkpoint helps you commit to the notes workflow you want to improve. A low-risk meeting, study, or planning use case gives your Action Brief Assistant a clear purpose by turning scattered details into reliable decisions, actions, owners, dates, questions, and missing information.

Set Up Your ChatGPT Project

Your chosen low-risk notes use case gives this assistant a clear job. A reliable brief also needs one place where its behavior can improve over time.

ChatGPT Projects keep related chats, files, and instructions together for recurring work. This dedicated workspace lets you improve one reusable workflow instead of starting over each time.

In this step, get ready to:

Reach the signed-in ChatGPT home screen in Safari.
Make a deliberate privacy decision using ChatGPT Data controls.
Create a tested ChatGPT Project named Action Brief Assistant.

Open ChatGPT on your Mac

The ChatGPT web app gives you access to Projects without installing extra software. Your account may ask for its usual verification step during sign-in.

Press Cmd+Space on your Mac to open Spotlight.
Type Safari into the Spotlight search field.
Press Enter to open Safari.
Select the address bar at the top of Safari.
Enter https://chatgpt.com in the address bar.
Press Enter to load ChatGPT.
Select Sign in if the landing page asks you to authenticate.
Complete the sign-in flow with your existing ChatGPT account.

You should see the signed-in ChatGPT home screen. Your account is ready for the project workspace.

Having Trouble Signing In?

Check that Safari has loaded the ChatGPT website before entering your account details.
Reload the page if the sign-in flow stops responding.
Help me troubleshoot ChatGPT sign-in on my Mac.

Review your privacy control

ChatGPT gives personal accounts a control for deciding whether new conversations can be used to train OpenAI models. Turning off Improve the model for everyone means new conversations are not used for that purpose.

Privacy settings can feel consequential, so take a moment to choose the option that fits your comfort level. The main project uses fictional notes regardless of your decision.

Select your account menu in ChatGPT.
Select Settings.
Select Data controls.
Review the current setting for Improve the model for everyone.
Decide whether new conversations can be used to improve OpenAI models.
Set the Improve the model for everyone toggle to match your decision.
Select Done to finish reviewing Data controls.

Your privacy choice is now deliberate. Keep using only the fictional notes supplied in the main project.

Create and test the project

The project gives your Action Brief Assistant a permanent home in the ChatGPT sidebar. For now, it stays free of reusable project instructions.

A short setup test confirms that the project chat accepts a precise request. This leaves you with a clean starting point for the experiments ahead.

Select New project in the ChatGPT sidebar.
Enter Action Brief Assistant in the project name field.
Select any icon for the project.
Select any color for the project.
Submit the project form.
Confirm that Action Brief Assistant appears in the sidebar.
Start a chat in the project workspace.

Before you send the test, do you expect the response to contain any extra explanation?

Test whether the project chat follows an exact instruction by sending this message:

Reply with exactly: Project ready.

What Does This Test Prove?

The message sets a precise output requirement. An exact reply confirms that the chat inside your project is active and following the request.

You should see only Project ready. in the response. That exact result completes the setup check.

Did ChatGPT Add Extra Text?

Confirm that Action Brief Assistant is selected in the sidebar.
Send the verification message again in a fresh chat inside the project.
Help me get the exact setup response in my ChatGPT Project.

Your Action Brief Assistant is live with a verified setup chat. Its reusable instructions remain empty for the first experiment.

Next, you'll run a basic summary prompt and expose the structure it leaves behind.

Test a Basic Summary

Your project in ChatGPT Projects is ready. The setup chat has already proved that it can respond inside the project.

A one-paragraph summary can sound useful while hiding the structure needed for action. This step gives you a baseline to compare with the stronger workflow you build next.

In this step, get ready to:

Run an intentionally limited summary prompt on the supplied webinar notes.
Compare the paragraph against the six sections needed for an action brief.
Record which sections are absent or mixed together.

Run the limited summary prompt

A baseline captures how a simple request behaves before you add reusable instructions. This prompt deliberately asks for one paragraph.

Start a new chat inside Action Brief Assistant from the project sidebar.

Before you send this, do you think one paragraph can give every decision, task, owner, date, question, and missing detail a clear home?

Submit the complete baseline prompt below in the message box:

Summarize the notes below in one paragraph.

Notes:
- Priya confirmed the onboarding webinar will move to November 12.
- Marco will update the registration page by November 8.
- We might add a customer Q&A, but nobody volunteered to own it.
- The budget is still waiting for finance approval.
- Lee asked whether the replay should remain available for 30 or 60 days.

What Does This Prompt Test?

The fictional notes include confirmed facts plus tentative ideas. They also include an unresolved approval plus an open question.

The one-paragraph constraint tests how much organization a basic summary provides on its own.

You should see a one-paragraph summary of the webinar notes. The six categories do not appear as distinct sections.

Seeing a Structured Response?

Confirm that you included the opening instruction limiting the response to one paragraph.
Start another new chat inside the project before resubmitting the complete prompt.
Help me check why my baseline prompt produced a structured response.

Compare the paragraph with the required brief

The six required headings form a simple output schema. Each heading gives one type of information a consistent place to live.

Before you inspect the format, do you expect every important detail in the paragraph to have a clear category?

Check whether Summary appears as its own section.
Check whether Confirmed Decisions separates firm choices.
Check whether Tentative Ideas isolates possible ideas.
Check whether Action Items shows an owner plus a due date.
Check whether Open Questions isolates unresolved questions.
Check whether Missing Information states what the notes omit.
Classify each required section as absent, mixed together, or distinct.

Your result should remain a paragraph instead of a six-section action brief. Some details may be present, yet their roles are absent or mixed together.

That baseline gives you concrete evidence of the structure your assistant needs. Next, you will turn this gap into reusable project instructions for every new chat.

Create a Structured Action Brief

Your baseline chat compressed the webinar notes into one paragraph. Its format blurred the boundaries between confirmed decisions and tentative ideas.

The ChatGPT Projects feature can apply reusable instructions to each fresh chat. Those rules enforce source grounding through a consistent six-section structure.

In this step, get ready to:

Configure reusable project instructions for structured action briefs.
Generate a fresh action brief from the original webinar notes.
Verify the structure and evidence-based classifications.

Add reusable project instructions

Project instructions define the assistant's standing job inside one project. They also set the evidence rules that every new project chat follows.

Select the project's more options menu (•••) in the Action Brief Assistant header.
Select Project settings.
Add the following text to the project instructions area:

You turn messy notes into a source-grounded action brief.

Use only information explicitly present in the notes.
Separate confirmed decisions from tentative ideas.
For every action item, list Task, Owner, Due date, and Evidence.
If an owner or due date is absent, write "Not stated".
Do not infer commitments from words such as might, maybe, could, or probably.

Use these sections in this order:
1. Summary, maximum 3 bullets
2. Confirmed Decisions
3. Tentative Ideas
4. Action Items
5. Open Questions
6. Missing Information

Keep the brief concise.

What do these instructions control?

The source rule limits every claim to information found in the supplied notes.
The certainty rule keeps tentative language out of confirmed commitments.
The action item structure places evidence beside every task.
The Not stated rule makes absent owners or dates explicit.
The ordered headings produce the same six-section shape in every brief.

Confirm the instructions area begins with You turn messy notes into a source-grounded action brief.
Confirm the final line reads Keep the brief concise.
Return to the Action Brief Assistant project chat.

Instructions missing or incomplete?

Confirm the project header shows Action Brief Assistant.
Return to Project settings from the project's more options menu.
Paste the complete instruction block again if any lines are missing.
Help me check why my ChatGPT project instructions are missing or incomplete.

Generate the brief in a fresh chat

A fresh chat tests whether the reusable instructions work without being copied into each request. Using the original notes also makes the new structure easy to compare with your baseline paragraph.

Start a new chat inside the Action Brief Assistant project.

Before you send the notes, do you think the reusable instructions will turn the same source into all six required sections?

Submit the original webinar notes by pasting this message:

Turn the notes below into an action brief.

Notes:
- Priya confirmed the onboarding webinar will move to November 12.
- Marco will update the registration page by November 8.
- We might add a customer Q&A, but nobody volunteered to own it.
- The budget is still waiting for finance approval.
- Lee asked whether the replay should remain available for 30 or 60 days.

What does this test prove?

The message supplies only the task and the original notes. The project instructions supply the evidence rules and output structure.

Reusing the baseline notes isolates the effect of your new instructions. A structured response shows that the reusable workflow is active.

You should see a structured action brief instead of another single paragraph.

Still seeing one paragraph?

Confirm the new chat appears inside the Action Brief Assistant project.
Return to Project settings to confirm the complete instructions remain visible.
Resubmit the exact notes after confirming the instructions.
Help me find why my project instructions did not structure the response.

Verify the six-section result

Before you inspect each section, do you expect the assistant to preserve every distinction from the source notes?

Confirm the headings follow this order: Summary, Confirmed Decisions, Tentative Ideas, Action Items, Open Questions, Missing Information.
Locate the webinar move to November 12 under Confirmed Decisions.
Inspect Marco's registration-page action for Owner Marco and Due date November 8.
Locate the customer Q&A under Tentative Ideas.
Locate the replay duration under Open Questions.
Confirm absent owners or dates display Not stated.
Confirm each action item includes supporting evidence from the notes.

You should see all six headings in the required order. Each classification should remain within the evidence supplied by the webinar notes.

That's the baseline gap closed. Your project now turns the same notes into a complete action brief without guessing missing details.

Your first source-grounded action brief is ready. Next, you'll challenge its certainty rules with ambiguous notes.

Audit Unsupported Certainty

Your first complete brief proved that reusable project instructions can impose a consistent structure. A polished response can still sound more certain than its source notes allow.

Now you'll pressure-test your ChatGPT assistant with ambiguous language. This puts source grounding under pressure before you depend on the brief.

In this step, get ready to:

Run an ambiguity stress test with confirmed statements plus tentative statements.
Compare every classification with its supporting note.
Ask the assistant to audit every claim for unsupported certainty.

Stress-test ambiguous notes

Tentative language can resemble a commitment when the response compresses several notes into a brief. This test combines possibilities with one explicit commitment.

Start a fresh chat inside the Action Brief Assistant project you already have open.

Before you send the notes, which statements do you expect the assistant to treat as tentative?

Submit the ambiguity stress test by sending this prompt:

Turn these notes into an action brief.

Notes:
- The team may release the beta on December 3.
- Nina could prepare the demo if design finishes the screens.
- Omar confirmed he will send the vendor shortlist by November 20.
- The budget might be $800.
- We still need someone to approve accessibility checks.

What does this prompt test?

The request relies on your saved project instructions to produce the same six-section structure.
The notes mix explicit confirmation with tentative wording. This tests whether the assistant preserves those evidence boundaries.

You should see an action brief with all six sections. Its classifications now give you evidence to inspect manually.

Missing the six-section brief?

Confirm that the new chat sits inside the Action Brief Assistant project.
Return to the project instructions from earlier.
Check that all six section names remain saved.
Help me understand why my ambiguity test did not use the saved project instructions.

Check the evidence boundaries

The assistant's formatting does not prove that its claims are grounded. Your manual comparison decides whether each classification matches the wording in the notes.

Keep the stress-test response visible in the current chat.
Compare each claim in the brief with the note that supports it.

Evidence Boundaries to Confirm

December 3 belongs with a tentative beta idea because the source says the team may release it then.
Nina's demo remains conditional because it depends on design finishing the screens.
Omar owns one confirmed action due November 20 because the source explicitly confirms his commitment.
The $800 budget remains tentative because the source says it might be that amount.
Accessibility approval remains missing assignment details. Its owner is Not stated. Its due date is Not stated.

Audit the assistant's claims

A self-audit creates a second pass through the evidence. Your manual review remains the final quality check.

Before you request the audit, do you expect the assistant to preserve every classification or revise an overstated claim?

Submit a claim-by-claim audit by sending this follow-up in the same chat:

Audit the brief against the notes. For each claim, identify its supporting note. Revise any claim that is stronger than the source.

How does the audit help?

The audit links each claim to the note that supports it. This makes unsupported wording easier to spot.
The revision instruction tells the assistant to weaken any claim that exceeds its evidence.

You should see supporting evidence for each claim. Any overstatement should appear in a revised brief.

Compare every supporting note with the assistant's audited claim.
Confirm that only Omar's vendor-shortlist commitment is treated as confirmed.
Confirm that the December 3 beta release remains tentative.
Confirm that Nina's demo remains conditional.
Confirm that the $800 budget remains tentative.
Verify that accessibility approval has its owner set to Not stated.
Verify that accessibility approval has its due date set to Not stated.
Use the revised brief as the corrected version if the audit identifies an overstatement.

Still seeing unsupported certainty?

Resend the audit prompt from above in the same ambiguity-test chat.
Check that the assistant identifies a supporting note for every claim.
Compare tentative words in the notes with the wording used in the revised brief.
Help me audit an action brief that still treats tentative notes as confirmed.

You handled the hardest judgment call here: the ambiguity chat now preserves both the stress test and its evidence audit.

Next, you'll test the assistant across several note styles to see whether this reliability holds consistently.

Evaluate and Strengthen the Assistant

Your ambiguity audit proved that the Action Brief Assistant can separate one confirmed commitment from several tentative claims. You have handled the hardest evidence boundary in this workflow.

One successful brief can happen by chance. A repeatable test set reveals whether the same behavior holds across different notes.

In this step, you will test three note sets inside your ChatGPT Project. You will use a manual scorecard to strengthen its reusable instructions.

In this step, get ready to:

Run three focused tests with complete information, missing details, or mixed certainty.
Score every result for structure, grounding, certainty, or missing fields.
Add a final quality-control rule before rerunning the weakest test.

Run three focused test cases

Each test isolates a different reliability risk. Separate chats keep each result tied to one source note set.

Start a fresh chat inside the Action Brief Assistant project.
Prepare Test A by pasting this prompt into the message box:

Turn these notes into an action brief.

Notes:
Rina approved the launch for January 15. Dev will send invitations by January 5.

What Does Test A Check?

Rina's approval tests whether an explicit decision remains confirmed.
Dev's invitation task tests whether the assistant preserves a named owner.
The January 5 deadline tests whether an explicit date reaches the action item.

Send the Test A message.
Compare the resulting brief with both source sentences.

You should see Rina's January 15 launch approval treated as confirmed. Dev should appear as the invitation owner with January 5 as the due date.

Is Test A Missing a Section?

Confirm the chat belongs to the Action Brief Assistant project.
Compare the response headings with the six headings in your project instructions.
Help me diagnose the missing section.

Start another fresh chat inside the Action Brief Assistant project.
Prepare Test B by pasting this prompt into the message box:

Turn these notes into an action brief.

Notes:
The team agreed to update the FAQ. No owner or deadline was assigned.

What Does Test B Check?

The agreed FAQ update tests whether confirmed work stays confirmed.
The missing assignment tests whether the assistant avoids inventing an owner.
The missing deadline tests whether the assistant avoids inventing a date.

Send the Test B message.
Compare the resulting brief with the source notes.

You should see the FAQ update treated as agreed work. The owner and due date should both display Not stated.

Did Test B Invent a Missing Detail?

Check whether the response assigns a person who never appears in the notes.
Check whether the response adds a deadline that never appears in the notes.
Help me find the unsupported detail.

Start a third fresh chat inside the Action Brief Assistant project.
Prepare Test C by pasting this prompt into the message box:

Turn these notes into an action brief.

Notes:
Sam might draft a demo. Jordan confirmed the budget is $1,200. Priya will book the room by February 2.

What Does Test C Check?

Sam's possible demo tests whether tentative language remains tentative.
Jordan's statement tests whether a confirmed budget remains confirmed.
Priya's room booking tests whether a clear commitment becomes an action item.

Send the Test C message.
Compare every claim with its supporting sentence.

You should see Sam's demo classified as tentative. Jordan's $1,200 budget should remain confirmed.

Priya should appear as the room-booking owner. February 2 should appear as the due date.

Did Test C Overstate Sam's Demo?

Find the section containing Sam's possible demo.
Confirm the wording preserves the uncertainty expressed by might.
Help me check the certainty levels.

Your three-case test set is complete. It now covers full details, missing information, or mixed certainty.

Score every result manually

A polished response can still contain unsupported claims. Your four-check scorecard turns reliability into evidence you can compare across every test.

Prepare the three score records by copying these lines:

Test A: Structure Pass | Grounding Pass | Certainty Pass | Missing fields Pass
Test B: Structure Pass | Grounding Pass | Certainty Pass | Missing fields Pass
Test C: Structure Pass | Grounding Pass | Certainty Pass | Missing fields Pass

How Does the Scorecard Work?

Structure passes when all six required sections appear.
Grounding passes when every fact exists in the supplied notes.
Certainty passes when confirmed statements stay separate from tentative statements.
Missing fields passes when every absent owner or date displays Not stated.

Replace a starter Pass with Fail whenever the response does not meet that check.

Return to the Test A chat from earlier.
Record Test A's four judgments with its matching score line.
Return to the Test B chat from earlier.
Record Test B's four judgments with its matching score line.
Return to the Test C chat from earlier.
Record Test C's four judgments with its matching score line.

The test with the most Fail ratings exposes the weakest behavior. Test C remains the strongest retest choice when every check passes because it mixes several certainty levels.

Add the final quality-control rule

The scorecard shows whether the assistant behaves consistently. A final source comparison gives every future brief one more chance to catch unsupported certainty.

Select the project's more options menu marked •••.
Select Project settings.
Locate the final instruction that says Keep the brief concise..
Place your cursor on a new line beneath that instruction.
Append the quality-control rule by pasting this sentence:

Before finalizing, compare every claim with the notes and revise any unsupported or overstated claim.

What Does This Rule Add?

The rule requires a final comparison between each claim and the source notes.
Unsupported claims must be revised before the brief is returned.
Overstated wording must be reduced to the certainty supported by the notes.

Confirm the new sentence appears as the final line of the project instructions.

✔️ Awesome, I've got everything!

Your reusable instructions now include the final source-comparison rule.

ⓧ I'd like to double check the full code

Compare your project instructions with this complete version.

You turn messy notes into a source-grounded action brief.

Use only information explicitly present in the notes.
Separate confirmed decisions from tentative ideas.
For every action item, list Task, Owner, Due date, and Evidence.
If an owner or due date is absent, write "Not stated".
Do not infer commitments from words such as might, maybe, could, or probably.

Use these sections in this order:
1. Summary, maximum 3 bullets
2. Confirmed Decisions
3. Tentative Ideas
4. Action Items
5. Open Questions
6. Missing Information

Keep the brief concise.

Before finalizing, compare every claim with the notes and revise any unsupported or overstated claim.

Return to the Action Brief Assistant chats from earlier.
Choose the test with the most Fail ratings.
Choose Test C if every test passed.
Start a fresh chat inside the Action Brief Assistant project.

Before you rerun the chosen test, consider this question: will the new rule change the brief or confirm that it already meets the standard.

Paste the same prompt with the same notes from your chosen test.
Send the message.

You should see all six sections in the required order. Every claim should remain within the evidence supplied by the chosen notes.

Apply the four scorecard checks to the rerun.

Is a Check Still Failing?

Return to Project settings to confirm the quality-control sentence remains the final instruction.
Repeat the test in a fresh project chat without changing its notes.
Help me diagnose the failed check.

Record the successful final result by pasting this line into the rerun chat:

Final rerun: Structure Pass | Grounding Pass | Certainty Pass | Missing fields Pass

What Does the Final Record Prove?

This line records that the refined instructions passed every evaluation check on a repeated test. The rerun demonstrates the improvement with evidence from a real output.

Send the final score line.
Confirm the rerun chat shows the action brief followed by four Pass results.

Strong finish. Your assistant now has three evaluation cases, a manual scorecard, or a final evidence check that protects every new brief.

Secret mission

Turn Your Own Notes into a Trusted Brief

Your assistant has passed fictional tests. Use a low-sensitivity note set to uncover one domain-specific weakness. The final proof is a four-check brief saved as a project source.

Clean Up Your Resources

This project costs $0 on ChatGPT Free. Choose whether to keep the project ready for future notes, leave it idle, or delete it permanently.

Resources you used:

The ChatGPT Project named Action Brief Assistant.

Keep everything running

No action needed. Choose this if you want to reuse the assistant for future low-sensitivity notes.

Reuse Action Brief Assistant from the sidebar when you have another low-sensitivity note set.
Continue refining the project instructions when your four-check scorecard reveals a specific weakness.

Pause - I'll come back to this later

Leave the project idle until you are ready to return. No teardown is required.

Leave Action Brief Assistant in the sidebar.
Close Safari when you are finished.

Delete - I don't want to use this again

Deleting the project is permanent. It removes its chats, instructions, and files stored only in the project.

Select the ••• menu beside Action Brief Assistant.
Select Delete project.
Complete the confirmation prompt to permanently delete the project.

You should no longer see Action Brief Assistant in the sidebar.

Files saved separately in Library remain available.

Nice Work!

You did it! Your reusable Action Brief Assistant now turns messy notes into structured briefs inside ChatGPT.

You've learned how to:

Build a reusable AI assistant that turns scattered notes into a consistent six-section action brief in every new project chat.
Apply source-grounded instructions that preserve certainty boundaries. Confirmed decisions stay distinct from tentative ideas. Absent owners or dates appear as Not stated.
Evaluate reliability through ambiguity stress tests. Your four-check scorecard verifies structure. It verifies grounding. It verifies certainty. It verifies missing fields.
Complete the optional Secret Mission with your own low-sensitivity notes. You audited every claim. You fixed a domain-specific weakness. You saved the trusted brief as a reusable project source.

Ready to quiz yourself?