Buiild it

Prompt

We are now officially entering the BUILD phase of Agency MarginGuard. I want you to act as a senior product architect, UX designer, workflow designer, AI-system designer, and practical digital-product builder. Do NOT merely describe the product. BUILD THE FIRST WORKING VERSION. Before building, briefly restate your understanding of the product in no more than 10 bullets. Then proceed directly into construction. ================================================== 1. PRODUCT DEFINITION ================================================== Product name: Agency MarginGuard Core promise: "Audit every closed deal for hidden scope, unpriced promises, and delivery risks before the kickoff." Target customer: Boutique digital agencies and AI automation agencies, approximately 3–20 employees, selling custom projects or retainers. Primary user: Agency owner, account manager, project manager, or delivery lead. Trigger: A deal has been marked Closed-Won and is being transferred from Sales to Delivery. Core problem: Important promises and expectations made during sales calls can differ from what is written in the SOW/proposal. Examples: - A salesperson promises additional deliverables. - A client is told revisions are flexible, but the SOW limits revisions. - A timeline is verbally promised without considering dependencies. - A service is discussed but never formally included. - Client assets/access are assumed but never collected. - The client expects something that the delivery team cannot find in the contract. The consequence is scope creep, unbilled work, delayed delivery, internal conflict, and client friction. ================================================== 2. V1 PRODUCT EXPERIENCE ================================================== V1 MUST WORK WITHOUT API INTEGRATIONS. The user manually provides: A. Sales-call transcript B. Signed SOW/proposal C. Optional onboarding/client information The system analyzes the materials and generates a structured: "MARGIN & HANDOFF AUDIT" The audit should contain: 1. Executive Summary 2. Overall Margin Leak Index (0–100) 3. Risk level 4. Critical findings 5. Verbal promises missing from the SOW 6. SOW contradictions 7. Ambiguous commitments 8. Timeline risks 9. Revision/change-request risks 10. Missing client dependencies/assets 11. Questions that must be resolved before kickoff 12. Recommended actions 13. Boundary-setting/change-order suggestions 14. Final kickoff readiness status ================================================== 3. EVIDENCE-FIRST ANALYSIS ================================================== This is one of the most important requirements. The AI must NEVER simply say: "Scope risk detected." It must show WHY. For every material finding, include: - Finding - Evidence from transcript/SOW - Source - Interpretation - Risk level - Recommended action Example: FINDING: LinkedIn profile optimization appears to have been promised verbally. TRANSCRIPT EVIDENCE: [Relevant short excerpt] SOW STATUS: No corresponding deliverable identified. INTERPRETATION: The client may reasonably expect this service. RISK: HIGH RECOMMENDED ACTION: Clarify whether the service is included. If not, resolve the expectation before kickoff. IMPORTANT: Do not invent quotations. If exact evidence cannot be located, explicitly say so. ================================================== 4. CONFIDENCE SYSTEM ================================================== Every important finding should have a confidence classification: CONFIRMED The evidence clearly supports the finding. PROBABLE Evidence strongly suggests the interpretation, but it is not explicit. AMBIGUOUS The material can reasonably be interpreted in multiple ways. NOT FOUND The system searched for corresponding evidence but did not find it. UNKNOWN The available material is insufficient to determine the answer. Do not turn uncertainty into certainty. ================================================== 5. MARGIN LEAK INDEX ================================================== Create a useful 0–100 scoring framework. The score should reflect risk rather than pretending to predict exact financial losses. Consider dimensions such as: - Unpriced promises - Scope ambiguity - Contract contradictions - Timeline commitments - Revision exposure - Missing client dependencies - Deliverable ambiguity - Approval/communication risk Design a transparent scoring methodology. The user must be able to understand WHY the project received its score. Do not use fake statistical precision. ================================================== 6. RISK CATEGORIES ================================================== Create a practical taxonomy. At minimum include: SCOPE PROMISES TIMELINE REVISIONS CLIENT DEPENDENCIES DELIVERABLES APPROVALS ASSUMPTIONS COMMUNICATION COMMERCIAL / PRICING Do not force a finding into a category if the evidence doesn't support it. ================================================== 7. HUMAN REVIEW ================================================== MarginGuard must not replace the account manager. The system should clearly separate: AI DETECTION from HUMAN DECISION. For example: AI DETECTION: "The transcript appears to contain a promise of three revision rounds." HUMAN DECISION: "Confirm whether this was intended to be included in the agreement." Never tell the user that the AI has made a legally binding determination. ================================================== 8. ACTION LAYER ================================================== The product must not stop at analysis. For every high-risk finding, provide a recommended next action. Examples: - Confirm with sales - Review SOW - Ask client for missing asset - Issue change order - Clarify expectation - Update project timeline - Document assumption - Escalate to agency owner ================================================== 9. CLIENT COMMUNICATION ================================================== Create professional templates for: A. Scope clarification B. Change-order request C. Missing client asset/access request D. Timeline dependency clarification E. Revision-limit clarification These should sound like real human agency communication. Do NOT write them like AI-generated corporate templates. Avoid phrases such as: "We're excited to optimize..." "To ensure maximum alignment..." "Leverage our collaborative ecosystem..." "Moving forward, we would like to..." Prefer direct human language. For example: "Before we start, I want to clarify one point from the sales call..." The communication should preserve the client relationship while protecting scope. ================================================== 10. HANDOFF WORKFLOW ================================================== Design the complete workflow: CLOSED-WON ↓ COLLECT MATERIALS ↓ RUN MARGIN GUARD AUDIT ↓ REVIEW FINDINGS ↓ RESOLVE HIGH-RISK ITEMS ↓ CONFIRM CLIENT DEPENDENCIES ↓ APPROVE HANDOFF ↓ KICKOFF Create a simple handoff checklist. ================================================== 11. PRODUCT ASSETS ================================================== Build the V1 product as a coherent package containing: 1. AI Audit Engine / prompt system 2. Margin Leak Index framework 3. Audit report structure 4. Handoff checklist 5. Boundary-setting templates 6. Change-order templates 7. User instructions 8. Example/demo case 9. QA/testing framework If a Notion workspace is practical, design the structure. If an interactive Artifact is the better V1 delivery mechanism, build a functional Artifact rather than merely describing one. ================================================== 12. DEMO CASE ================================================== Create a realistic fictional agency scenario to demonstrate the system. Create: - fictional sales transcript - fictional SOW - fictional onboarding information Then run MarginGuard against them. The demo should contain realistic contradictions and hidden promises. The output should demonstrate the value of the product. Do NOT use real people's names or claim that fictional findings are real-world statistics. ================================================== 13. AI-SLOP RESISTANCE ================================================== This requirement is NON-NEGOTIABLE. The final product must feel like it was designed by someone who understands agency operations. It should NOT feel like: "ChatGPT + prompts + a Notion template." Avoid: - excessive emojis - fake urgency - excessive bolding - generic business buzzwords - repetitive explanations - unnecessary AI references - giant walls of text - overly polished marketing language - fake case studies - fabricated statistics - meaningless "insights" Use concrete examples. Use realistic operational language. Make the interface and reports calm, professional, and useful. The user should be able to imagine opening this after a client signs a $10,000 project. ================================================== 14. PRODUCT QUALITY BAR ================================================== Before declaring the build complete, test it against the fictional demo case. Ask: 1. Did it catch the important hidden promises? 2. Did it avoid false positives? 3. Can an account manager understand the result quickly? 4. Does every major finding have evidence? 5. Does the score make sense? 6. Does it tell the user what to do next? 7. Do the client emails sound human? 8. Could an agency actually use this tomorrow? 9. Does it feel substantially more useful than simply pasting the transcript into ChatGPT? 10. Is there unnecessary complexity that should be removed? If something fails, improve it before declaring V1 complete. ================================================== 15. IMPORTANT LIMITATIONS ================================================== Do not claim: - proven $8,700 average savings - guaranteed ROI - guaranteed elimination of scope creep - legal accuracy - universal compatibility - that the product has been validated by real agencies These are hypotheses or future validation targets. The current objective is: BUILD A HIGH-QUALITY V1 THAT IS READY FOR REAL-WORLD TESTING. ================================================== 16. OUTPUT FORMAT ================================================== Do not spend the entire response explaining what you COULD build. Build it. Where possible, create usable artifacts/files/interfaces rather than giving me instructions to create them myself. At the end, provide a concise: BUILD STATUS with: - What was built - What works - What remains - What I should test first - Any technical limitation caused by the free Claude environment Do not recommend Claude Pro unless a specific limitation genuinely prevents the current build. START NOW.

A system prompt was added to support web rendering