Khurram Badar / Archive / Papers / BRIEF 3: CONSUMER LAYER — Block Reader / Pedestrian Judgment App

BRIEF 3: CONSUMER LAYER — Block Reader / Pedestrian Judgment App

briefing · 2026-05-09 · 1178 words · Khurram Badar

BRIEF 3: CONSUMER LAYER — Block Reader / Pedestrian Judgment App Purpose Build the first user-facing agent that solves the problem Khurram lived last night: he searched "tikka biryani" on Google Maps, the right place (Pak Liyari) showed up, but the pin sent him to.

ai · content

BRIEF 3: CONSUMER LAYER — Block Reader / Pedestrian Judgment App

Purpose

The product: a conversational, mobile-first AI agent that *answers the actual question* a pedestrian is asking — including the things Google Maps doesn't know (back entrances, real prices, what's actually open, what people actually order).

Project Name

Strategic context (read first)

Prerequisite

Tech stack

The single core user flow

1. User opens the web app on phone
2. Sees a prompt: *"What are you looking for?"*
3. Types or speaks: *"chicken tikka, cheap, walking distance, now"*
4. App calls our AI with: query + user's GPS + base layer DB context for nearby shops
5. AI returns ONE recommendation as a structured response:
- **Place name**
- **Distance + walk time**
- **What to order** (extracted from reviews via base layer enrichment)
- **What it costs** (best estimate, marked as estimate)
- **The thing Google Maps doesn't tell you** (back entrance, alley access, payment quirks)
- **A "Walk me there" button** that deep-links to Google Maps with destination pre-loaded
6. User can tap follow-ups: *"too far"*, *"anything closer"*, *"is it busy now"*, *"do they take card"*
7. AI answers each without rebuilding the whole result

The "AI prompt" that powers the recommendation

When the user query comes in, send Claude API a prompt like:
```
You are a local Bur Dubai friend. Answer the user's question with ONE place
recommendation, not a list. Use the structured shop data provided. Be specific
about: the door to enter, what to order, the price range, and any quirks Google
Maps wouldn't tell them (back entrance, payment method, busy hours).

User location: <lat,lng>
User query: <text>
Nearby shop data (from base layer): <JSON>
Time: <current time>

Constraints:
- One recommendation, not five
- Be specific about price ranges (use AED, mark as estimate)
- If you don't know, say "I don't know" — do NOT guess
- If multiple shops fit, pick the one with strongest evidence and explain WHY
- Keep response under 80 words
- Never recommend a place we have no photo evidence for
```

The user interface (mobile-first)

Screen 1: Input

Screen 2: Result

Screen 3: Follow-up chat

The Pak Liyari demo case (must work end-to-end)

This is the gold-standard test:
- User stands at Grand Astoria
- Types: "chicken tikka or biryani, cheap, near me"
- App returns:
> **Pak Liyari Restaurant Br.** — 50m, ~2 min walk. The mutton biryani is what 5,000+ reviews recommend; ask for "double masala" to lift the rice. ~18-25 AED for a full plate. **Tip:** the entrance is on 21A Street, *behind* your hotel — not the front. Walk around the south side. Cash preferred.
- "Walk me there" button opens Google Maps with Pak Liyari pre-loaded

If this works, the product works.

The Google Maps deep-link

Honest constraints

Output structure

Repo

Deployment

What this is NOT

Tone of the AI responses (critical)

Success criteria

Future (not for v1, just noting)

← Underwater Dive Simulation — Claude Code Build PromptBRIEF 1: BASE LAYER — Structured Semantic Map of Zone 312, Bur Dubai →
Two years of working thought, indexed.
Ask me to present it in your conference room — WhatsApp +971 55 623 9111
Book Session →