AI Lead Triage Lab with Claude

Build a reusable Claude workflow to triage leads with human review.

Introduction

30 Second Summary

A confident reply can still be wrong when it fills in details that nobody provided. For customer inquiries, that confidence can turn missing information into a risky promise.

In this project, you will build an AI Lead Triage Lab in Claude that turns fictional customer inquiries into evidence-based follow-up drafts for human review. The finished workflow keeps missing details visible for your final judgment.

What You'll Build

Your finished project turns a fictional inquiry into the same review-ready triage format in every fresh chat.

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

A baseline comparison that shows where a vague prompt chooses its own labels or assumptions.
An evidence-based triage table with five consistent fields. Each priority points back to the inquiry. Missing facts appear as Unknown or Needs Review.
A reusable Claude Project that applies your tested rules in a fresh chat. A four-check rubric keeps the final judgment with you.
Secret Mission: Add a human escalation rule for unsupported urgency and guarantee requests.

Do I need coding experience?

A Claude account is all you need. The whole project runs in your web browser with fictional inquiries.

Before We Start

Before the hands-on work begins, commit to building a fictional AI lead-triage workflow in Claude. A human reviews every priority decision and draft because Claude can make unsupported assumptions or produce confident but unreliable results.

Set Up a Safe Claude Practice Project

Your human-review commitment now needs a clean workspace for the lead-triage exercise. A Claude Project keeps this practice separate from unrelated chats.

The supplied fictional inquiries give you enough material to practise safely. Real customer information stays outside the lab.

In this step, get ready to:

Sign in to Claude on the web.
Create the AI Lead Triage Lab project.
Keep the project restricted to supplied fictional inquiries.

Sign in to Claude on the web

For this exercise, Claude runs in your browser. Your existing account is all you need.

Open your web browser.
Go to Claude Projects.
Sign in with the Claude account you already have.

You should land on the Projects page. Your Claude workspace is ready without any local setup.

Having Trouble Signing In?

Check that your browser has internet access. Confirm that you are using the email address linked to your Claude account.

If the sign-in flow still stops, ask for help with your Claude sign-in issue.

Create the practice project

A dedicated project gives this exercise a recognisable home. Its description records the purpose of the workspace before you add any practice inquiries.

Click + New Project in the upper-right corner of the Projects page.
Enter AI Lead Triage Lab as the project name.
Enter this project description: Practice workspace for learning prompt design and lead triage.
Submit the completed project form.

You should see the new project page with AI Lead Triage Lab as its name. Your named practice space now separates this exercise from unrelated chats.

Why Use a Separate Project?

The project groups this lab in its own workspace. That boundary makes the later prompt comparisons easier to follow.

A clear name also helps you return to the correct workspace before testing another fictional inquiry.

Project Not Created?

Return to the Projects page. Check whether AI Lead Triage Lab appears in your project list.

If the project is missing, repeat the form with the exact name and description above. Ask for help with the project creation screen.

Set the safety boundary

Customer messages can contain information that belongs outside a practice exercise. This lab uses fictional inquiries so you can focus on prompting without exposing real people or business records.

Use only the supplied fictional inquiries in this project.
Keep real names out of the project.
Keep real phone numbers out of the project.
Keep financial details out of the project.
Keep credentials out of the project.
Keep health records out of the project.
Keep confidential business documents out of the project.
Leave the project free of real customer files.

Before you check, what should a safe practice project contain at this point?

Confirm that the project page shows AI Lead Triage Lab.
Check that the project contains no customer files.
Check that the project contains no customer messages.

You should see AI Lead Triage Lab on the project page. The project should contain no real customer data, files, or messages.

Your safe practice workspace is ready for fictional lead inquiries. Next, you will give Claude a vague request and inspect the assumptions it makes.

Test a Vague Prompt

Your safe workspace in Claude is ready for fictional data. Now you need a baseline that shows how the model responds before you define success.

A large language model can produce a polished answer from a vague request. This step lets you test that answer before deciding which instructions are missing.

In this step, get ready to:

Send three fictional inquiries through one deliberately vague request.
Inspect the response for decisions Claude made without guidance.
Capture the baseline for comparison with your next prompt.

Build the baseline message

A baseline gives you something concrete to compare with later results. Keeping the inquiry data unchanged isolates the effect of your instructions.

Start a new chat inside the open AI Lead Triage Lab project.
Paste these three fictional inquiries into the message composer:

Lead A: Our office cooling stopped this morning and the rooms are getting hot. Can someone visit today?
Lead B: Please send information about routine maintenance options for next month.
Lead C: We need help. Please contact us.

What Are These Inputs Testing?

Lead A reports a present-day operational problem.
Lead B describes a future information request.
Lead C provides almost no detail.

Check that the composer shows Lead A through Lead C in one unsent message.

Missing Part of an Inquiry?

Delete the partial inquiry text. Copy the complete block again so every word stays consistent across your tests.

Help me compare my pasted fictional inquiries with the supplied baseline inputs.

Place your cursor after Lead C.
Add one blank line.
Add the baseline request by copying this line below the inquiries:

Prioritize these leads and write replies.

What Does This Request Do?

This request names two broad tasks. It gives Claude freedom to decide what priority means.

It also leaves the response structure open. The result shows which choices the model makes for itself.

Before you send the message, do you think all three leads will receive equally well-supported decisions?

Send the completed message.

You should receive priorities or replies for Leads A, B, and C. This is your first visible baseline.

No Baseline Response?

Confirm that the three inquiries and the baseline request were sent as one message. If the response stopped early, ask Claude to continue its current answer without adding new instructions.

Help me understand why my baseline lead-triage response is missing or incomplete.

Inspect Claude's choices

A baseline becomes useful when you separate the inquiry evidence from the choices Claude supplied. Your review tests whether each decision has support in the source text.

Before you inspect the response line by line, which part do you think gave Claude the most room to choose its own approach?

Review the priority labels Claude chose.
Compare every claim with the words in its inquiry.
Compare the structure of the three draft replies.
Flag wording whose certainty exceeds the inquiry evidence.

You should find that Claude chose labels without receiving a definition of priority. Depending on the response, you may also find unsupported assumptions.

You may also find different reply structures. Some wording may sound more certain than the inquiries justify.

You've captured the baseline we needed. Next, you'll turn its gaps into a clear AI brief.

Create a Clear Triage Brief

Your baseline prompt gave you a plausible result to inspect. It also showed Claude filling undefined gaps with its own choices.

A clear brief fixes that gap by defining success before Claude responds. Each priority now has a meaning tied to the inquiry text.

In this step, get ready to:

Define four priority labels with evidence-based criteria.
Build a structured prompt with constraints and a fixed output format.
Test the brief against the original fictional leads.

Define the priority criteria

A priority label is useful when another person can trace it back to a rule. These four criteria create that trail.

What does each priority mean?

Use High when the inquiry explicitly reports an immediate operational or safety impact. Use it when a service is unavailable now.
Use Medium when the inquiry describes a clear service need without an immediate operational impact.
Use Low when the inquiry is informational or exploratory. Use it for work planned for the future.
Use Needs Review when the inquiry lacks enough evidence for another label.

Return to the baseline response in the project chat from earlier.
Check whether every baseline priority maps to one of the four definitions above.

Build the structured brief

A strong brief separates the model's role, task, context, constraints, and output format. This gives Claude a bounded job with a visible standard for completion.

Click inside the message box beneath the baseline response.
Paste this complete brief without submitting it yet:

Role: lead-triage assistant for a local service business. Task: classify each inquiry and draft a short follow-up. Priority criteria: High = explicit immediate operational or safety impact, or a service unavailable now; Medium = clear service need without immediate operational impact; Low = informational, exploratory, or planned future work; Needs Review = insufficient evidence for another label. Context: respond quickly without inventing customer details or commitments. Constraints: use only inquiry text; mark missing facts as 'Unknown'; never invent a price, appointment, diagnosis, or guarantee. Output format: a table with Lead, Priority, Evidence, Missing information, and Draft reply.

How does this brief control the result?

The Role gives Claude one business perspective for the task.
The Task limits the work to classification and a short follow-up.
The Priority criteria connect every label to observable evidence.
The Context explains why speed must stay grounded in the inquiry.
The Constraints protect missing facts with Unknown and block unsupported commitments.
The Output format places every result into the same five-column table.

Before you submit the brief, which lead do you expect to receive the least confident label?

Submit the message in the same project chat.

You should see a table containing one row for each of the three fictional leads. The table should use the five fields named in your brief.

That is the key shift: your brief now controls what useful means.

Missing a column or seeing invented details?

Check that you pasted the entire brief. The final sentence must name all five columns.

If Claude still invents a detail, submit the complete brief again. Keep the original inquiries unchanged so the comparison stays fair.

Help me troubleshoot my structured triage response.

Compare the structured result

The inquiries stayed the same between both runs. The new result reveals what changed because of your instructions.

Before you compare the responses, which improvement do you expect to notice first?

Scroll to the baseline response from earlier.
Compare its priority labels with the structured table.
Check the Evidence column for wording quoted from each inquiry.
Check the Missing information column for Unknown where facts are absent.
Inspect the Draft reply column for unsupported prices or appointments.
Inspect the same drafts for unsupported diagnoses or guarantees.
Confirm that Lead C uses Needs Review when its text cannot support another label.

Your structured response should show Lead, Priority, Evidence, Missing information, and Draft reply as columns. It should quote each inquiry as evidence.

You should also see Unknown or Needs Review wherever the inquiry cannot support a firmer answer. A human remains responsible for every priority and draft.

You have closed the vague-prompt gap. Every triage decision now has rules a human can inspect.

Your criteria now make the original leads easier to review. Next, you will use worked examples to show Claude what careful evidence and cautious language look like.

Add Examples and Edge Cases

Your structured prompt now gives Claude explicit priority labels. Your previous test proved that those instructions produce a five-column answer grounded in the inquiry text.

Written rules can still leave tone or borderline cases open to interpretation. Worked examples show Claude the safe pattern to follow.

In this step, get ready to:

Add a worked High example with quoted evidence.
Add a worked Needs Review example that preserves unknown facts.
Test the updated prompt with a borderline inquiry.

Add two worked examples

A worked example pairs one fictional inquiry with an acceptable model answer. Its explanation connects the answer to your evidence rules.

Return to the message composer beneath the structured result in your current project chat.
Replace the previous prompt with this example-enhanced version:

Role: lead-triage assistant for a local service business.
Task: classify each inquiry and draft a short follow-up.
Context: the business wants to respond quickly without inventing customer details or commitments.
Priority criteria:
- High: explicit immediate operational or safety impact, or a service unavailable now.
- Medium: clear service need without immediate operational impact.
- Low: informational, exploratory, or planned future work.
- Needs Review: insufficient evidence for another label.
Constraints: use only the inquiry text. Mark missing facts as 'Unknown'. Never invent a price, appointment, diagnosis, or guarantee.
Output format: a table with Lead, Priority, Evidence, Missing information, and Draft reply.

Worked High example:
Inquiry: Our office cooling stopped this morning and the rooms are getting hot. Can someone visit today?
Example output:
| Lead | Priority | Evidence | Missing information | Draft reply |
| Lead A | High | "cooling stopped this morning" and "rooms are getting hot" | Preferred visit time: Unknown. Site access details: Unknown. | Thank you for letting us know. What time works best today? We will review availability before confirming a visit. |
Why acceptable: The quoted text supports High. Missing facts remain Unknown. The draft requests scheduling details without promising a visit.

Worked Needs Review example:
Inquiry: We need help. Please contact us.
Example output:
| Lead | Priority | Evidence | Missing information | Draft reply |
| Lead C | Needs Review | "We need help" does not identify a service need or its urgency. | Urgency: Unknown. Service type: Unknown. Budget: Unknown. Timing: Unknown. | Thank you for contacting us. What service do you need? How urgent is it? What timing do you have in mind? Do you have a budget range? |
Why acceptable: The quoted text does not support another priority. Missing facts remain Unknown. The draft asks focused questions without making a commitment.

How the examples guide Claude

The High example connects an immediate outage to quoted evidence.
The Needs Review example demonstrates caution when the inquiry lacks usable facts.
The explanations show why each reply stays within your constraints.

Add the original fictional inquiries beneath the example-enhanced prompt:

Lead A: Our office cooling stopped this morning and the rooms are getting hot. Can someone visit today?
Lead B: Please send information about routine maintenance options for next month.
Lead C: We need help. Please contact us.

Why reuse the same inquiries?

The inquiry data stays constant during this test. Any change in the result now comes from the worked examples.

Send the combined prompt and inquiries to Claude.

You should see another table with Lead, Priority, Evidence, Missing information, and Draft reply. Each row should follow the patterns demonstrated by the examples.

That upgrade is now working. Claude has concrete demonstrations of grounded evidence and cautious replies.

Examples not shaping the result?

Check that both examples appear above the three inquiries in the same message. Confirm that each example includes all five output columns.

Make sure each explanation appears directly beneath its matching example.

Ask for help with the response you received.

Inspect the example-guided result

Examples only help when they demonstrate every behavior you care about. Check the new response against both patterns before testing a harder case.

Locate the row for Lead A.
Check that its Evidence cell quotes the present-day outage.
Check that its Missing information cell preserves unavailable scheduling details as Unknown.
Check that its Draft reply avoids promising a visit.
Locate the row for Lead C.
Check that its priority reflects insufficient evidence.
Check that its missing-information field keeps urgency, service type, budget, and timing unknown.
Check that its reply asks focused follow-up questions.

What makes these examples acceptable?

Each priority traces back to words in the inquiry. The Evidence field makes that connection visible.

Unavailable facts stay Unknown. This prevents a plausible answer from quietly becoming an invented fact.

Each draft asks for what the business needs next. It avoids unsupported commitments.

Test a borderline inquiry

An edge case probes the boundary between two labels. This inquiry reports a problem while explicitly saying the system still works.

Paste this fictional inquiry into the same project chat:

The system is making a strange noise, but it still works. We would like someone this week.

What does this edge case test?

The inquiry combines a current symptom with a functioning system. It tests whether Claude can preserve that distinction in its evidence and reply.

Before you send it, do you think Claude will preserve the difference between a strange noise and a failed system?

Send the edge-case inquiry to Claude.

You should see the same five-column format. The response should quote the strange noise while avoiding any claim that the system has failed.

That is the edge-case win: your prompt keeps the response tied to what the customer actually said.

Compare the Evidence field with the evidence patterns in your worked examples.
Compare the Missing information field with the unknown-fact patterns in your worked examples.
Compare the Draft reply with the cautious language in your worked examples.

Seeing an unsupported failure claim?

Check that the constraint limiting evidence to the inquiry text remains in your updated prompt. Confirm that the prompt still forbids invented diagnoses.

Ask for help identifying the instruction Claude overlooked.

✔️ My edge case stayed grounded

Your result uses the five requested columns. It also preserves the fact that the system still works.

ⓧ I'd like to double check the full prompt

Compare your example-enhanced prompt with this complete version.

Role: lead-triage assistant for a local service business.
Task: classify each inquiry and draft a short follow-up.
Context: the business wants to respond quickly without inventing customer details or commitments.
Priority criteria:
- High: explicit immediate operational or safety impact, or a service unavailable now.
- Medium: clear service need without immediate operational impact.
- Low: informational, exploratory, or planned future work.
- Needs Review: insufficient evidence for another label.
Constraints: use only the inquiry text. Mark missing facts as 'Unknown'. Never invent a price, appointment, diagnosis, or guarantee.
Output format: a table with Lead, Priority, Evidence, Missing information, and Draft reply.

Worked High example:
Inquiry: Our office cooling stopped this morning and the rooms are getting hot. Can someone visit today?
Example output:
| Lead | Priority | Evidence | Missing information | Draft reply |
| Lead A | High | "cooling stopped this morning" and "rooms are getting hot" | Preferred visit time: Unknown. Site access details: Unknown. | Thank you for letting us know. What time works best today? We will review availability before confirming a visit. |
Why acceptable: The quoted text supports High. Missing facts remain Unknown. The draft requests scheduling details without promising a visit.

Worked Needs Review example:
Inquiry: We need help. Please contact us.
Example output:
| Lead | Priority | Evidence | Missing information | Draft reply |
| Lead C | Needs Review | "We need help" does not identify a service need or its urgency. | Urgency: Unknown. Service type: Unknown. Budget: Unknown. Timing: Unknown. | Thank you for contacting us. What service do you need? How urgent is it? What timing do you have in mind? Do you have a budget range? |
Why acceptable: The quoted text does not support another priority. Missing facts remain Unknown. The draft asks focused questions without making a commitment.

Your prompt now demonstrates how to handle clear evidence and ambiguous requests. Next, you will evaluate the workflow with observable checks before saving it for fresh project chats.

Evaluate and Save the Workflow

Your example-enhanced prompt now handles a borderline inquiry without inventing a system failure.

A polished Claude response can still be wrong. Human review catches unsupported claims before a draft reaches a customer.

A human evaluation rubric gives you a repeatable quality gate.

Saved project instructions remove the need to repaste the workflow in every new project chat.

In this step, get ready to:

Apply four human-review checks to the three original leads.
Save the tested workflow as project instructions.
Verify the saved instructions in a fresh project chat.

Evaluate the original leads

An evaluation rubric turns quality into observable checks. The inquiry remains the source of evidence for your final judgment.

Return to the project chat containing your example-enhanced prompt.
Prepare the rerun by pasting these fictional inquiries beneath the tested prompt:

Lead A: Our office cooling stopped this morning and the rooms are getting hot. Can someone visit today?
Lead B: Please send information about routine maintenance options for next month.
Lead C: We need help. Please contact us.

Why Rerun the Same Inquiries?

This rerun keeps the inquiry data unchanged. The evaluation isolates the behavior created by your instructions.

Before you send the inquiries, which rubric check do you expect to be hardest for Claude to pass?

Send the message in Claude.

You should see a new triage output covering all three leads.

Your Four-Check Human Rubric

Every requested field is present.
The priority is supported by words in the inquiry.
Missing details are marked Unknown instead of invented.
The reply is concise and professional. It contains no unsupported price, diagnosis, appointment, or guarantee.

Mark each check Pass or Fail for Lead A.
Mark each check Pass or Fail for Lead B.
Mark each check Pass or Fail for Lead C.

Keep the Human Decision Final

Claude can critique its own output. That critique remains model-generated input.

Use its critique as another perspective. Keep your own Pass or Fail marks as the final decision.

How to Target a Failed Check

A missing field points to the five-column output instruction.
An unsupported priority points to the evidence rule.
An invented detail points to the Unknown requirement or the prohibition on invented facts.
A weak draft points to the reply constraint.

Revise one connected instruction if any check is marked Fail.
Rerun all three original leads after making a revision.
Reapply the four-check rubric to the revised output.

Good work. Your workflow now has a human quality gate that checks evidence before confidence.

Save the tested project instructions

Project instructions apply saved behavior to every chat inside the same Claude Project. This gives your tested workflow a reusable home.

Why Project Instructions?

A normal chat holds the prompt inside one conversation. Project instructions carry the tested behavior into fresh chats within AI Lead Triage Lab.

What the Saved Instructions Must Include

Include the tested lead-triage role and classification task.
Include the High, Medium, Low, and Needs Review criteria.
Include the synthetic-data safety boundary and inquiry-text evidence rule.
Include the Unknown requirement and the prohibitions on invented prices, appointments, diagnoses, or guarantees.
Include the Lead, Priority, Evidence, Missing information, and Draft reply columns.
Include the worked High and Needs Review examples with their explanations.

Select the tested prompt from its role through its worked-example explanations.
Copy the selected prompt.
Return to the AI Lead Triage Lab project page.
Select Set project instructions.
Paste the tested prompt into the project instructions field.
Add the synthetic-data safety boundary from your project setup.
Review the instruction field against the checklist above.
Select Save instructions.

Your tested workflow is now saved at the project level. Fresh project chats can use it without another copy of the long prompt.

Instructions Not Saving?

Confirm you are editing project instructions inside AI Lead Triage Lab. Check that the instruction field contains text before selecting Save instructions.

Help me check why my Claude Project instructions are not saving.

Verify the workflow in a fresh chat

A fresh chat tests whether the saved instructions carry the workflow by themselves. This check removes the earlier conversation as a source of hidden context.

Start a fresh chat from inside the AI Lead Triage Lab project.
Confirm the project name appears on the fresh chat page.
Prepare the final check by pasting this fictional inquiry:

We are comparing maintenance plans for later this year.

What Does This Test Prove?

The fresh chat contains only one synthetic inquiry. A structured response therefore shows that the project instructions supplied the role, criteria, constraints, examples, and output format.

Before you send the inquiry, do you think Claude will use the five-column format without seeing the long prompt in this chat?

Send the inquiry in Claude.

You should see a table containing Lead, Priority, Evidence, Missing information, and Draft reply without repasting the full prompt.

Apply the four-check rubric to the fresh response.
Keep your human review as the final decision on its priority and draft.

Missing the Five-Column Format?

Confirm the fresh chat belongs to AI Lead Triage Lab. A chat outside the project cannot use these project instructions.

Return to Set project instructions if a required column is missing. Confirm the tested output format is present before selecting Save instructions again.

Help me find why my fresh Claude Project chat is ignoring the saved format.

You have turned a one-off prompt into a reusable Claude workflow. Every fresh result still passes through human review before it is used.

Secret mission

Add a Human Escalation Rule

A request can sound urgent while hiding every fact needed for a safe decision. Add an escalation rule for unsupported urgency, contradictory information, and guarantee requests. Add a Human action field that tells the reviewer what to confirm.

Clean Up Your Resources

Choose whether to keep, archive, or delete your Claude Project. Claude Free costs $0, so this project has no ongoing cost.

Resources you used:

Claude Project named AI Lead Triage Lab.
Saved project instructions containing the reusable lead-triage workflow.
Synthetic project chat history containing the baseline, structured tests, edge case, fresh-chat verification, and escalation test.

Keep everything running

No action needed. Choose this if you want to practise with more synthetic inquiries or refine the workflow.

Your AI Lead Triage Lab project remains active in Claude.
Your saved project instructions continue to guide fresh project chats.
Your synthetic chat history remains available for comparison.
Continue using only synthetic inquiries to preserve the project's safety boundary.

Pause - I'll come back to this later

Archive the project to pause your practice while keeping its conversations available.

Select the ... menu in the open AI Lead Triage Lab project.
Choose the archive action.

Your archived project conversations remain accessible when you return.

Delete - I don't want to use this again

Remove all project resources and start fresh. Deleting the project removes its saved workflow and synthetic chat history.

Select the ... menu in the open AI Lead Triage Lab project.
Select Delete.
Select Yes, delete. to confirm.
Return to Projects.
Confirm that AI Lead Triage Lab is no longer listed.

Nice Work!

Well done! Your reusable AI lead-triage workflow in Claude now turns synthetic inquiries into evidence-based priorities for human review.

You've learned how to:

Created a visible baseline that exposed how a vague prompt can invent its own priorities.
Built a structured triage brief that classifies fictional inquiries using quoted evidence. Missing facts stay Unknown.
Applied a four-check human evaluation rubric to original leads and edge cases. Saved project instructions now reproduce the tested output in a fresh chat.
Secret Mission: added a human escalation rule that routes unsupported urgency, contradictory information, and guarantee requests to Needs Review. The added Human action field tells the reviewer what to confirm before replying.

Ready to quiz yourself?