Claude Code Build Night · Part 7 · Coming soon

Growth Engineering

Find where users leak out, and fix one thing at a time.

How the night runs.

A short AI news briefing, then we instrument What's for Dinner live and find its biggest leak, then five-minute demos from the room.

  1. 01

    Define

    Five stages, written for What's for Dinner.

  2. 02

    Measure

    Real events and one dashboard.

  3. 03

    Fix

    Name the biggest leak, ship one experiment.

  4. 04

    Nudge

    One email for people about to drift away.

The deck.

Every slide from the night, free to share. Open it full size to step through.

Growth Engineering · slidesOpen full size →

Tonight's metrics plan.

The growth plan we write on the night: five stages defined for one app, the events behind them, the dashboard, one experiment and one email. Copy it and swap in your own app.

GROWTH.mdWhat's for Dinner

What's for Dinner

Type what's in your fridge. Get three dinners you can make tonight.

The goalTurn signups into people who cook with the app every week.

  • PostHog
  • Resend
  • Neon
  • Next.js
  • Vercel
Done when
  • Five stages defined for this app
  • signed_up and dinners_found arriving, tied to user ids
  • A dashboard with funnel, weekly actives and week-4 retention
  • The biggest leak named, one experiment running
  • The at-risk email working, with address and unsubscribe
The full metrics plan
# GROWTH.md · What's for Dinner

## Goal
Turn signups into people who cook with the app every week. Find the biggest leak, fix one thing at a time.

## The five stages
| Stage | Definition for What's for Dinner | Event or query |
|---|---|---|
| Acquisition | A new account is created | `signed_up` (with utm_source, utm_content) |
| Activation | Their first three dinners found, within 24 hours of signup | first `dinners_found` |
| Engagement | Finds dinners in a given week | `dinners_found` at least once that week |
| Retention | Still finding dinners in week 4 after signup | week-4 cohort retention |
| Churn | No `dinners_found` for 14 days | days since last `dinners_found` >= 14 |

At risk: 10 to 13 days since the last `dinners_found`.

## Events
- `signed_up`: server-side, after the account is created. Properties: utm_source, utm_medium, utm_campaign, utm_content
- `dinners_found`: server-side, after three recipes come back. Properties: ingredient_count, seconds
- `dinners_failed`: server-side, when the AI call fails or times out
- `dinner_opened`: client-side, when someone opens a recipe's steps

Identify users by their user id, never by email. No ingredient lists or recipe text in event properties.

## Dashboard
1. Funnel: signed_up -> first dinners_found (activation rate)
2. Weekly active users: unique users with dinners_found per week (engagement)
3. Retention: weekly cohorts, signed_up -> dinners_found, through week 4
4. At risk and churned: counts from the dinners table
5. Activation by utm_content: which creators bring people who cook

## Biggest leak
The stage with the lowest conversion from the stage before it. Write it down with the number before choosing a fix.

## Experiment
One change, one metric, one decision date.
- Hypothesis: if the empty fridge input shows three tappable starter ingredients, more new users reach their first dinners
- Metric: activation rate
- Split: 50/50 feature flag on new signups
- Run: until each variant has at least 100 signups, or two weeks
- Decide: keep, kill or rerun. Log it here either way.

## Lifecycle email
"What's in your fridge tonight?" sent once to at-risk users (10 days quiet), at most once every 30 days per person.
- Sent with Resend from our verified domain
- One link back to the app with utm_source=email&utm_campaign=at-risk
- An unsubscribe link that works in one click, plus List-Unsubscribe and List-Unsubscribe-Post headers, honored before every send
- Our physical postal address in the footer: [your Stable address] from Part 1
- Never sent to anyone who unsubscribed or never activated
- Advisor question, not legal advice: if any users are in the EU, UK or Canada, do their consent rules mean we need an opt-in before sending this?

## Weekly review
Every Monday: the five stage numbers, the change from last week, the experiment status, one next step.

## Done when
- [ ] The five stages defined in this file
- [ ] signed_up and dinners_found arriving in analytics, tied to user ids
- [ ] A dashboard with funnel, weekly actives and week-4 retention
- [ ] The biggest leak named with its number, and one experiment running
- [ ] The at-risk email sent to a test account, with a postal address and a working one-click unsubscribe

The prompts.

Every prompt from the deck, in order. Open one, copy it, paste it into Claude Code, Codex or any coding agent.

  1. 01
    Write the definitions first
    Read SPEC.md and GTM-PLAN.md. Help me write GROWTH.md for this app.
    
    We use five stages: acquisition, activation, engagement, retention and churn. Ask me one question at a time to define each one for this app, as a specific event or query, not a vague goal. Suggest a default for each and let me change it. Push back if a definition can't be measured.
    
    Then write GROWTH.md with: Goal, The five stages (a table of stage, definition, event or query), Events (name, when it fires, properties, server or client), Dashboard, Biggest leak, Experiment, Lifecycle email, Weekly review, Done when.
    
    No personal data or ingredient text in event properties. Stop and wait for my OK.
  2. 02
    Instrument the events
    Read GROWTH.md.
    
    1. Tell me to sign up for PostHog and run npx @posthog/wizard in my own terminal. Wait until I tell you it's done, then read what it changed and summarize it.
    2. Add server-side capture with posthog-node for the server events in GROWTH.md (signed_up, dinners_found, dinners_failed). Use the user id as the distinct id, call identify on the client after sign-in, and flush events before the route returns.
    3. Add the client events in GROWTH.md, and nothing else custom.
    4. Never log email addresses, ingredient lists or recipe text as properties.
    5. Tell me to sign up a test account and find dinners once, then confirm each event arrived with the right properties. Stop there.
  3. 03
    One dashboard
    Read GROWTH.md.
    
    1. Tell me to add PostHog's MCP server by running npx @posthog/wizard mcp add in my own terminal, then restart Claude Code and log in when asked. Wait until I tell you it's done.
    2. Use the PostHog tools to create a dashboard called "What's for Dinner growth" with the insights listed under Dashboard in GROWTH.md: the activation funnel, weekly active users, weekly retention through week 4, and activation by utm_content.
    3. For at-risk and churned counts, write a small SQL query against our dinners table and tell me how to run it.
    4. Give me the dashboard link and today's number for each insight. If there isn't enough data yet, say so plainly.
  4. 04
    Find the biggest leak
    Read GROWTH.md and use the PostHog tools and our database.
    
    1. For the last 30 days, give me the conversion from each stage to the next as a table, with the number of users at each step.
    2. Name the stage with the biggest drop. If the sample is too small to trust, say how many users we need first.
    3. For that stage, look at what users did before dropping off (session recordings, if I've turned them on, and events) and list three possible causes, each with the evidence.
    4. Write the leak and its number under Biggest leak in GROWTH.md.
    
    Don't change any code. Stop and wait.
  5. 05
    Ship one experiment
    Read GROWTH.md, especially Biggest leak and Experiment.
    
    1. Restate the hypothesis, the one metric, the split and the decision rule in four lines. Wait for my OK.
    2. Use the PostHog tools to create the experiment and its feature flag as a draft, and show me what you made. Wait for my OK before starting it.
    3. Build the variant behind the flag, for example three tappable starter ingredients on an empty fridge input. The control stays exactly as it is. New signups only.
    4. Show me both variants at phone width. Log the experiment, start date and decision date in GROWTH.md, then stop.
  6. 06
    Catch them before they churn
    Read GROWTH.md, section Lifecycle email.
    
    1. Tell me to create a Resend account, verify our sending domain and add RESEND_API_KEY to .env.local myself. Wait until I tell you it's done. Never print the key.
    2. Add an email_opt_out column and a one-click unsubscribe route that sets it.
    3. Write a route that finds activated users with no dinners for 10 to 13 days who haven't opted out and weren't emailed in the last 30 days, then sends "What's in your fridge tonight?" with resend.emails.send: two sentences, one tracked link back to the app, the unsubscribe link, and a footer with our physical postal address (use the placeholder [your Stable address] from Part 1 and tell me to fill it in). Add List-Unsubscribe and List-Unsubscribe-Post: List-Unsubscribe=One-Click headers that point at the same route, so mail apps can unsubscribe in one click. Record each send.
    4. Add an advisor question to the Lifecycle email section of GROWTH.md: if any users are in the EU, UK or Canada, do their consent rules mean we need an opt-in first? It's a checklist item, not legal advice.
    5. Protect the route with CRON_SECRET and add a daily Vercel cron for it.
    6. Run it once against a test account only, and tell me to check the inbox, the postal address footer and the unsubscribe link. Stop there.
  7. 07
    Review it every Monday
    Read GROWTH.md. Using the PostHog tools and our database, write this week's review:
    1. The five stage numbers, and the change from last week.
    2. Experiment status: users per variant, the metric for each, and whether we've reached the decision rule. Don't call a winner early.
    3. Lifecycle email: sent, clicks back to the app, unsubscribes.
    4. One next step.
    
    Append it to a Weekly review log in GROWTH.md and stop.
  8. 08
    Retention by source
    Bonus round. Read GROWTH.md and GTM-PLAN.md.
    
    Using the PostHog tools or our database, show activation and week-4 retention grouped by utm_content, for sources with at least 10 signups. Sort by week-4 retention, not signups. Flag any group too small to trust.
    
    Then suggest one change to the lead scoring in GTM-PLAN.md based on which kinds of creators bring people who stay. Wait for my OK before editing it.

The toolkit.

The tools that measure the habit and nudge people back, one job each.

Count what matters. Fix one leak.

Date to be announced. Dates go out in the Discord first.

Harbor at dusk · Phoenix, Arizona