Build Sprint · Day 11 · Prompt Pack

Multi-modal in 120 minutes.

3 paste-ready prompts. Take your text-only MVP and add one more sense. Fresh signals from Super AI 2026 baked in.

🌏 Super AI signal woven into every prompt. The "Heard at Super AI" callouts share what Megan brought back from Suntec + the NEXT AI Hackathon — embedded directly into the prompt logic, not just decoration.
⚡ How to read these prompts · 60 seconds
Each prompt has a specific real persona panel in its ROLE section. The CONSTRAINTS section is non-negotiable. The SELF-CHECK forces measurable rules. Run them in order: P00 (optional) → P01 (pick) → P02 (build) → P03 (audit). 40 min total.
🔁 Two rules for the WHOLE pack
1. Context reuse. Claude already knows your PRD, MVP cut, user journey, Lovable URL from Day 7. If you're in the same chat, skip the brackets it knows.

2. Save every output as .md in your project folder. After each prompt, ask: "Save this as `[name].md` in my project folder."
Quick jump · 3 expert-grade prompts + auto-context
Phase 0 · Optional first move
Skip the typing.
Prompt 00 · Auto-Context
Let Claude fill the brackets for you.

Optional. Run ONCE at the top of your chat. Claude introspects what it knows about your project (Day 7 outputs, Lovable URL, MVP cut) and confirms before auto-filling brackets in P01–P03.

Optional · First
Paste at the top of your chat · once
# ROLE You are my Build Sprint multi-modal scope partner for the next 40 minutes. You'll help me pick the right modality pair, wire it into Lovable, and audit cost. # GOAL Before we start: read what you know about my project from prior context — Day 7 PRD, MVP cut (1-3 features), user journey, Lovable URL, QA report, any cost work. Then confirm with me. # OUTPUT Restate, in this exact format: - **Product (one line):** [your understanding] - **Target user (specific):** [your understanding] - **The 1-3 MVP features:** [your understanding] - **Live Lovable URL:** [your understanding] - **Current modalities used:** [likely "text only"] - **My scariest modality (if known):** [your understanding] End with: "Reply 'yes' if accurate, or correct any line. I'll auto-fill these in P01–P03 for the rest of the session." # DELIVERABLE After confirmation, tell me: "Save this confirmed context as `multimodal_context.md` in my project folder." # SELF-CHECK - ✓ Did I list ALL 6 fields above? - ✓ Did I use the user's actual words where possible? - ✓ Am I asking for confirmation, not assuming?
Phase 1 · Pick the pair
3 candidates → 1 winner.
Prompt 01 · Modality Pair Picker
Three pairs. One winner. Picked by panel.

A panel of three senior multi-modal product designers generates 3 modality-pair candidates for YOUR product · scores each on shipability/wow/cost/journey-fit · picks the winner with reasoning · gives you the v1.5 architecture.

Pick
Run after Prompt 00 (or solo) · 5 min
# ROLE You are a panel of three senior multi-modal product designers who've shipped at scale at top consumer AI companies. You think in causal chains: which modality creates the strongest user moment? Which adds the most cost? Which can a Build Sprint cohort ship in 48 hours? # GOAL For my product, generate 3 different modality-pair candidates. Score each on 4 axes. Pick the winner with reasoning. # CONTEXT (Claude reuses prior context — fill only what's needed) NOTE · Claude already knows your PRD, MVP cut, user journey, Lovable URL from Day 7. Fill what's new. - The 1-3 MVP features (from Day 7 Prompt 02): [FEATURES OR "USE PROJECT CONTEXT"] - Target user: [USER OR "USE PROJECT CONTEXT"] - Live Lovable URL: [URL] - Modalities currently used (likely just text): [E.G. "TEXT ONLY"] - Days to ship the multi-modal upgrade: 2 (deadline = tomorrow's hackathon demo) - Modalities I'm scared of: [FROM THE EARLIER HANDS-UP] # CONSTRAINTS - 3 pair candidates · each pair is exactly 2 modalities (not 3, not 1) - One pair MUST include my scariest modality (we don't avoid it) - Each scored on 4 axes (1–5): shipability in 2 days · wow factor · added cost · journey fit - The winner is picked by panel consensus, not highest score # SUPER AI CONTEXT — bake this into your reasoning At Super AI 2026, the cohort heard 3 signals worth weighing: 1. Voice is the new mobile — sub-150ms TTFB · real revenue · the modality with biggest moat right now 2. Video gen cost dropped ~40% in 6 months — if you wrote off video for cost, re-evaluate 3. Image editing models (Flux Kontext class) are the next image wave — worth scoring as "image" for B2C products When scoring "wow factor", account for these shifts. # OUTPUT FORMAT **🎯 The 3 Pair Candidates** **Pair A · [Modality 1] + [Modality 2]** - The user experience this enables (1 sentence) - The "wow per second" moment (what they see in 0–10 sec) - Shipability: __/5 · Wow: __/5 · Cost: __/5 · Journey fit: __/5 **Pair B · [Modality 1] + [Modality 2]** - Same structure **Pair C · [Modality 1] + [Modality 2]** (MUST include my scariest modality) - Same structure **🏆 The Winner** The panel picks: Pair [A/B/C] - Reasoning (3 sentences max) - One thing each panelist would change about the runner-up **⚡ The v1.5 architecture (winner only)** - Step 1 (LLM does): [action] - Step 2 ([modality 2] does): [action] - Step 3 (user sees): [the share moment] **🚨 Tomorrow's risks** - 3 specific failure modes (latency · cost spike · API failure · UI confusion) - < 30 sec recovery for each # DELIVERABLE Tell Claude: "Save this as `modality_pair.md` in my project folder." # SELF-CHECK - ✓ Exactly 2 modalities in each pair (not 3, not 1)? - ✓ Pair C includes my scariest modality? - ✓ Winner buildable in 2 days? - ✓ Did I factor in the Super AI signals when scoring wow? - ✓ Did I (Claude) save `modality_pair.md`?
⚡ Pro tip · sit with the discomfort The winning pair is rarely the one you guessed in the pair exercise. Founders pick what flatters their original vision. The panel picks what ships fastest with the most wow. Trust the panel for v1.5. You can pivot for v2.
Super AI signal · what this prompt encodes
Voice scoring bias. If your panel doesn't score voice highly for a product where speed-of-interaction matters, push back. After Super AI 2026, "voice is the new mobile" is the consensus — Cartesia + ElevenLabs Turbo both hit sub-150ms. Voice is no longer a wildcard.
Phase 2 · Wire it
From spec → live Lovable app.
Prompt 02 · Lovable Multi-Modal Wire-up
Paste once. 2 modalities wired live.

Paste-ready Lovable prompt. Takes your chosen modality pair → updates your 3-screen MVP · adds Supabase generations + rate_limit tables · sets up API key handling via edge functions · publishes. Built-in fallback for APIs Lovable can't connect to directly.

Build
Paste directly into Lovable · existing project · 15 min
# ROLE You are a senior Lovable + multi-modal engineer who has wired ElevenLabs, Flux, Luma, and HeyGen APIs into 100+ Lovable apps. You build on the first try. You name the cheapest fitting model. You wire rate-limits before launch, not after the surprise bill. # GOAL Take my existing Lovable MVP (3 screens · text-only · auth via Supabase) and add my chosen modality pair. Output: paste-ready Lovable prompt I can drop in immediately. # CONTEXT (Claude reuses prior context) - Existing Lovable URL: [URL OR "USE P05 OUTPUT"] - Chosen modality pair (from Prompt 01): [E.G. "TEXT + VOICE" OR "USE modality_pair.md"] - For each new modality, the tool I want to use: [E.G. "VOICE = ELEVENLABS · IMAGE = FLUX"] - API keys I have: [E.G. "CLAUDE, FLUX. NEED: ELEVENLABS"] - Free-tier signup tolerance: [E.G. "OK TO ADD 1 NEW VENDOR ACCOUNT"] # SUPABASE SCHEMA UPDATE — DO THIS FIRST Add to my existing Supabase project: generations table (tracks usage + cost) - id uuid primary key - user_id uuid references auth.users - modality text (e.g., 'image', 'voice', 'video') - vendor text (e.g., 'flux', 'elevenlabs') - cost_cents integer - inputs jsonb - output_url text - created_at timestamptz default now() rate_limit table (enforces per-user caps) - user_id uuid primary key - generations_today integer default 0 - last_reset_at timestamptz default now() Add RLS: users read their own rows · cannot write to rate_limit (writes happen via edge function only). # BUILD FLOW — exactly these steps Screen 2 update (the Working screen) - Add UI for the new modality's input (file upload · text · voice record button) - On submit: check rate_limit table · if over cap, show friendly "you're at your daily limit" - Otherwise: call the modality API · insert row in generations table with cost - Update rate_limit counter Screen 3 update (the Aha moment) - Display the new modality output prominently (image · audio player · video player) - "Download" button + "Copy share link" button (the viral loop) API key handling - Store all API keys in Supabase Edge Function secrets (NEVER in frontend) - Each modality call goes through an edge function (e.g., /generate-image, /generate-voice) - Edge function checks rate limit · makes the call · inserts cost row · returns output URL # DESIGN RULES (non-negotiable) - Mobile-first (390px baseline) - Light theme · accent matching original landing page - Loading state on every modality call (these take 2–30 sec) - Error state with specific message (not "Something went wrong") # PUBLISH 1. Wait for Lovable build to complete 2. Test in Lovable preview · submit a test action 3. If Lovable adds a 4th screen → reply: "Reduce to 3 screens. Add modality to Screen 2." 4. Click Publish (top-right rocket) 5. Wait 10-15 sec for live URL 6. Test on phone · confirm cost shows in Supabase generations table 7. Save live URL as `live_url_v15.txt` # DELIVERABLE Tell Claude: "Save this build summary as `multimodal_wireup.md` in my project folder." Include: live URL, new schema, API keys needed, any fallback edge functions created. # SELF-CHECK — answer YES before "Done" - ✓ Are modality calls going through edge functions (not frontend keys)? - ✓ Is rate-limit enforced before each call? - ✓ Is cost inserting into generations table? - ✓ Does the Aha screen show the output + share button? - ✓ Is it published to a live URL? - ✓ Did I resist building a 4th screen? - ✓ Did I (Claude) save `multimodal_wireup.md`? End with: live URL + Supabase dashboard link.
⚠️ If Lovable can't handle direct integration · edge function pattern Some modality APIs (Luma video, ElevenLabs streaming, HeyGen avatars) don't have clean Lovable connectors. Use this pattern instead:
  1. Build the modality call as a Supabase Edge Function (Deno runtime)
  2. Lovable frontend calls it via supabase.functions.invoke('generate-voice', { body })
  3. Edge function: validates auth · checks rate limit · makes API call · inserts cost row · returns output URL
  4. Lovable frontend displays the returned URL in Screen 3
If a vendor isn't supported, this pattern always works. Slower to build (~5 extra min) but bulletproof.
Super AI signal · prompt caching = 80% savings
Anthropic's Super AI session was clear: if you're not caching system prompts and tool definitions on long-context Claude calls, you're overpaying 5x. After this Lovable wire-up, add a one-liner to your edge function: cache_control: { type: "ephemeral" } on stable parts of your prompt. This alone will move your cost line in Prompt 03.
Phase 3 · Audit the burn
Know the cost before the demo.
Prompt 03 · Cost & Latency Audit
$X at 100 users. $XX,XXX at 10K.

For your newly multi-modal app, Claude projects costs at 100 / 1K / 10K users. Names 3 specific cost-saving moves. Gives you a free-tier strategy that keeps the first 100 users under $50. Includes a hard kill-switch if monthly spend exceeds your budget ceiling.

Audit
Run AFTER Lovable build is live · 5 min
# ROLE You are a senior ML infrastructure engineer who has shipped to millions of users at major AI labs. You think in cost per call · cost at scale · cost of failure. You also have a finance-aware PM brain — cofounders fold not when the product fails but when the bill hits. # GOAL For my newly multi-modal app, project costs at 100 / 1,000 / 10,000 monthly users. Name 3 specific cost-saving moves. Give me a free-tier strategy that lets me onboard the first 100 without burning $200. # CONTEXT (Claude reuses prior context) - My multi-modal stack (from P01 + P02): [E.G. "CLAUDE SONNET + ELEVENLABS + SUPABASE" OR "USE multimodal_wireup.md"] - Average actions per user per month: [E.G. "10 ACTIONS/MO IN MVP"] - Each action calls these APIs: [E.G. "1 LLM CALL + 1 IMAGE GEN"] - My current free tier limits per user: [E.G. "3 ACTIONS/DAY"] - My monthly budget burn ceiling (above which I pause growth): [E.G. "$500"] # CONSTRAINTS - Project costs at EXACTLY 100, 1K, 10K users - Cost-saving moves must be specific implementations (not "use a smaller model") - Free-tier strategy must keep first 100 users under $50 total - Honest about which moves are real vs theoretical # SUPER AI CONTEXT — factor these into your projections - Prompt caching (Anthropic): 80%+ savings on long-context calls. If we're not using it, recommend it as Move A. - Cartesia vs ElevenLabs: Cartesia is 5x cheaper at near-equal quality for non-cloned voices. If we're using ElevenLabs for non-emotional use, recommend swap as Move B. - Open-source multimodal small models (Llama 4 vision, DeepSeek): catching up at 1/10 cost. Worth mentioning as a v2 consideration. # OUTPUT FORMAT **💸 Cost Projection · Monthly** | Users | LLM | Image | Voice | Video | Total | Per-user | |-------|-----|-------|-------|-------|-------|----------| | 100 | $X | $X | $X | $X | $X | $X | | 1,000 | ... | ... | ... | ... | ... | ... | | 10,000| ... | ... | ... | ... | ... | ... | **📊 The breakdown** - 80% of cost comes from: [specific modality + call type] - The single most expensive call per user: [specific] - The model swap that saves the most: [specific recommendation] **🔥 The 3 cost-saving moves** 1. **Move A** — [specific cache key strategy OR specific cheaper model OR specific tier change] 2. **Move B** — [specific change] 3. **Move C** — [specific change] **🆓 Free-tier strategy** - First 100 users: limit to [N actions/week] - Estimated total spend at 100 users: $X - Convert to paid tier when: [specific trigger] - Paid tier offering: $[X]/mo for unlimited **🚨 The kill-switch** If monthly spend exceeds my $500 budget, I should: 1. [specific action 1 — e.g., disable video generation tier] 2. [specific action 2 — e.g., move all new signups to waitlist] 3. [specific action 3 — e.g., switch ElevenLabs → Cartesia] # DELIVERABLE Tell Claude: "Save this as `cost_audit.md` in my project folder." Reference it every Monday before deciding to scale or hold. # SELF-CHECK - ✓ Did I project costs at exactly 100 / 1K / 10K? - ✓ Are the 3 cost moves SPECIFIC to this stack? - ✓ Does the free-tier strategy keep first 100 under $50? - ✓ Is there a hard kill-switch with 3 named actions? - ✓ Did I factor in the Super AI signals (caching, Cartesia, OS models)? - ✓ Did I (Claude) save `cost_audit.md`?
⚡ Pro tip · the 1K column is the truth-teller Ignore the 100-user column. Stare at the 1K column. That's where most builders' jaws drop. If 1K-user monthly burn is > 30% of your budget ceiling, ship Move A this weekend. The cohort that audits cost before launch is the cohort that survives a viral moment.
Super AI signal · the cost line that surprised me
At Super AI, I heard 5 different builders share that they hit $5K+ monthly bills within 2 weeks of launch because they shipped without rate limits. Voice was the #1 culprit (ElevenLabs streaming on a viral product). This prompt is the prevention. Don't skip it.
Copied to clipboard