# R1 Product and build lead: attack on MVP-PLAN v2 ## 1. Verdict The idea is testable, but v2 builds a product (full dashboard, attribution, call tracking, reports) to test a question that is mostly about ads and dealer behaviour. Cut the build to a car page, a grounded agent and SMS, run everything else by hand, and fix the sold-flow before you write dashboard code. ## 2. Top 5 problems **1. The dashboard is the biggest build and the biggest bet, and it fails on the plan's own make-or-break question.** Independent lot owners on a phone, on the lot, in the evening, will not log in to a new app to "mark sold". Attribution, stock freshness and the CPA all depend on that habit. Fix: make SMS the dealer interface for the pilot. A daily text: "Still have the 2016 CR-V? Reply Y / SOLD / HOLD". "SOLD" gets a follow-up: "Who bought it? 1) Dana (asked Tue) 2) Sam 3) Someone else". Magic links, no passwords. Build a minimal web form for add-car (phone-first, VIN, km, price, photos) and nothing else. Leads and bookings go to a Google Sheet or Airtable view. Build the real dashboard only after the pilot shows what they actually touch. **2. Week 2 is overloaded and has hidden calendar waits.** Real inventory, SMS, alerts, "arrived", who-bought-it, call tracking, weekly report, a lot audit, a new Meta Page and ad account, all in 5 days, plus signing a dealer. Meta ad account, Business verification and ad review can take days (ASSUMPTION, based on common experience, not measured). A Canadian SMS number may need toll-free verification, which is also slow (ASSUMPTION; check the provider's current docs on day 1). Fix: start the waiting items on day 1 of week 1: buy a Canadian number and submit its verification, create the Meta Business Manager, and use a Page under our own business first. Cut call tracking, the weekly report build and the buyer-check automation from week 2 (see scope). **3. Agent reliability: do not let the model write facts.** A "never makes up price" instruction in a prompt is not a control. Fix: the model only picks tools and writes friendly text. Price, km, availability and slots are rendered from database fields by code (a car card), never typed by the model. Tools: `search_cars`, `get_car`, `get_slots`, `book_slot`, `handoff`. Post-check: any number or dollar figure in the reply must match a field in the tool result, or the reply is replaced by a handoff. Stale stock: a car not confirmed in 48 h shows "ask to confirm" and takes no bookings. Build a 40-question eval set (price tricks, "is it still there", "will you take $X", finance, trade-in, French, rude, off-topic) and run it before every deploy. Handoff must be a rule, not a judgement: any of negotiate, finance, trade value, warranty, "call me", or two low-confidence turns hands off with a text to the dealer and a promise to the buyer ("Mike will text you by 6 pm"). During the pilot a team member reads every conversation live and can take over. **4. SMS and Canadian numbers.** Buyers on a chat widget give a number; unsolicited or after-hours texting from an unregistered number gets filtered (ASSUMPTION). CASL consent and STOP handling must be built, not noted. Fix: use a Canadian local or toll-free number from Twilio or similar, one sender per dealer, store consent text and timestamp with each lead, honour STOP in code, and keep texts transactional (booking code, reminder, one follow-up). No marketing blasts in the pilot. Send nothing between 9 pm and 8 am local. **5. Attribution is too clever for a pilot with one dealer and low volume.** Seven steps, a 60-day any-car rule and an automated buyer survey are built before a single sale exists. Expected volume is a handful of sales (ASSUMPTION). Fix: log lead, booking code and phone number (that is the whole attribution engine). Do the "who bought it" and the buyer check by hand: every Friday, 10 minutes with the dealer on the phone using a sheet of our leads, and we text 10-day buyers ourselves. Automate only after 20+ leads. Also: the plan's own gap, that Marketplace is free, remains unanswered. The pilot must measure the share of leads that are new buyers (not people already in the dealer's DMs) via a "how did you find us" field. ## 3. Scope **Must build (real)** - Public car page with all-in price, photos, specs, chat box and slot booking - Grounded agent with the tools above, the eval set, and full transcript logging - SMS: Twilio-style Canadian number, booking confirmation, dealer alerts, STOP, consent log - Add / hold / sold by SMS plus one mobile add-car form - Leads, bookings, sold, who-bought-it in one Postgres (Supabase) **Fake or manual** - Dashboard: Supabase table editor or an Airtable interface, shared with the dealer read-only - Lot audit: done by us by hand; photos uploaded by us. Spend the half-day once, but keep it to the top 15 cars, not all 80 - Weekly report: a Friday email written by hand from a sheet - Ads: created and managed by us manually in Meta; no campaign tool - Buyer check, review flags, "arrived" tracking: by phone or text, by hand - Call tracking: one forwarding number on the ads only, no recording; or cut it - Trade-in: one free-text question captured, no valuation **Cut** VIN-decoder automation (type year, make and model by hand; a free decoder is optional later, and ASSUMPTION: coverage of Canadian-market vehicles needs checking), Google Search test, Instagram, per-dealer branding, auto-flagging, any financing work, multi-dealer features. **Realistic days** (stack: Next.js on Vercel, Supabase, Twilio, one LLM API; one full-stack person plus one doing ads and dealer contact) - **W1 D1:** schema, seed 8 demo cars, car page. Order the SMS number and open Meta Business. - **W1 D2:** agent with tools, card rendering, 40-question eval. - **W1 D3:** chat widget, slot booking, add-car form on mobile. - **W1 D4:** demo polish, 90-second screen recording as backup, pitch and discovery sheet. - **W1 D5:** dealer visits. Hit the discovery questions hardest on mark-sold and Marketplace. - **W2 D6:** SMS in and out, dealer alerts, STOP and consent. Sign the pilot dealer (a go on dealer interest is already a result). - **W2 D7:** lot audit, 15 cars loaded. Ads submitted for review. - **W2 D8:** sold-by-text flow and "who bought it". Morning stale-stock text. - **W2 D9:** team acts as 20 fake buyers, fix the eval failures, check privacy page and consent copy. - **W2 D10:** buffer. Go live Monday of week 3. ## 4. What to keep Wedge C with A inside it. Stock-matched car pages. Booking over lead capture. The "mark sold asks who bought it" idea (move it to SMS). Meta first. Pilot without fee, then monthly, then CPA only after the numbers match. All-in pricing and no financing quotes. A kill criterion on dashboard use. ## 5. One missing product idea **Text-the-car intake.** The dealer texts or WhatsApps photos and a price to our number. The agent reads the plate or VIN photo, asks for km, and creates the listing and car page in two minutes. This replaces both the lot audit (cost to serve, which does not scale at half a day per dealer) and the app login problem, and it keeps stock fresh because adding and selling a car are the same texting habit.