Take apart a successful service's growth with evidence, then rebuild it.
Pick one of Jumia·YouTube·Jiji, observe it, reverse-engineer it into 3 docs, build with AI, then prove how it differs from the original. It's not copying screens — it's a project applying growth principles to a new context.
First visit · 30 minWhile logged out, record the value, choices, CTAs and your 5-second impression.
Main task · 45 minRecord clicks, scrolls and inputs, expected vs actual results, and trust/drop-off factors.
Return features · 30 minFind where saving, subscribing, notifications, reviews and sharing lead to the next visit or value for others.
03
Link evidence with IDs
OBSERVATION TEMPLATE
## OBS-001 — [observation title]\n- Date / URL / environment:\n- User goal:\n- Fact observed:\n- My interpretation:\n- Effect on the next action:\n- AARRR stage:\n- Evidence file:\n- Uncertainty:
Observation: "Product cards show ratings and review counts." / Interpretation: "This seems to reduce uncertainty before opening details." Don't mix the two.
04
Turn AARRR into a repeating growth loop
User action→Service value grows→Other users respond→Acquisition·return
Write the 3 strongest OBS, the hidden links, the conditions where the loop stops, the principles to borrow and the things you must not copy.
PASS GATE
Mission 1 done check
0 / 8
MISSION 02 · DEFINE
Reverse-Engineer Your Observations into 3 Docs
Turn Mission 1's evidence into actionable contracts: a product opportunity, a business model and a base engineering design.
25POINTS
Input OBS · GROWTH LOOPOutput 3 core MDsDone signal OBS→PRD→BM→ADR
01
Start with an evidence trace table
Trace in one line where each decision started. This is the central tool that stops the docs from describing different products.
Can it be demoed without real personal data, payments or outside services?
Login, payment, live chat and recommendation algorithms can be replaced with mock data.
04
Give AI contradiction checks, not generation
CROSS-DOC REVIEW PROMPT
The three docs below define the product, business and tech of the same service.\nDon't create new features or facts; find only these:\n1. PRD/BM/ADR items with no OBS evidence\n2. Places using the same term differently\n3. Conflicts between product scope and dev scope\n4. Success criteria that can't be measured\n5. Personal data, copyright and security risks
PASS GATE
Mission 2 done check
0 / 8
MISSION 03 · BUILD
Build the Site with AI Based on Your Docs
Use the 3 docs as a build contract to make a main user flow with start, key action, done and failure states.
30POINTS
Input PRD · BM · ADROutput WEB · TEST · AI LOGDone signal LINT·BUILD pass
01
Check the environment and build baseline first
node --version npm --version git --version
npx create-next-app@latest development/apps/web
After installing, record the build target, linked IDs, how to confirm done and out-of-scope items in `development/README.md`. If the official Next.js requirements differ from this guide, follow the official docs and note why in an ADR.
02
Finish in 3 vertical cycles
Understanding the valueMake it possible to understand the audience and value and find the CTA within 5 seconds.
Key actionState changes through browsing or input, handling loading, empty results and wrong input.
Done and next actionLink done feedback to the growth hypothesis's next action, like saving, sharing or browsing more.
03
Lock the scope and done criteria for every AI request
AI BUILD PROMPT
Role: Next.js development partner\nGoal: [user result for this cycle]\nDocs you must follow: PRD-___ / BM-___ / ADR-___\nPaths you may change: [files or folders]\nDone criteria: screen conditions / failure states / mobile·accessibility / tests\nDon'ts: new features, copying original assets, secrets, unverified packages\nFirst suggest a work plan and the files to change.
After each task, record the request, changed files, checks, AI errors and human change/reject decisions in `AI-###`.
04
Leave reproducible checks, not just what you saw
FeaturesNormal·empty·error·done states
AccessibilityKeyboard·focus·WCAG
Responsive360px / 1440px
Codenpm run lint / build
PerformanceLighthouse before·after
TestsReproduce the main flow or E2E
PASS GATE
Mission 3 done check
0 / 8
MISSION 04 · LEARN
Prove the Differences Between the Original and Your Build
Use the same task and conditions on both sites to prove where growth principles were kept, improved, lost or changed.
20POINTS
Input original·build·TESTOutput CMP 10+ · EXP 3Done signal evidence for every conclusion
01
Fix the same comparison baseline
First record the original and build versions, measurement time, device, browser, screen size, login state and the shared task. Compare "how each reaches the same goal," not "are they similar."
Same start→Same task→Same end→Same conditions
02
Repeat 3 times, switching the order
Alternate between the original and your build, recording completion, time, clicks and inputs, blockers and errors. Also write the limit that testing alone doesn't represent all users.
Understanding the value5-second record
Task efficiencyTime·steps·errors
Growth structureNext action·return
TrustEvidence of less worry
BM linkValue exchange
Tech qualityAccessibility·performance
OriginalityAdapting the principle
03
Make CMP records your basis for judgment
COMPARISON TEMPLATE
## CMP-001 — [comparison title]\n- Area / linked evidence:\n- What you saw in the original:\n- What you saw in your build:\n- Category: kept/improved/lost/intentional change\n- Evidence file or measurement:\n- Why the difference happened:\n- Effect on users:
04
Close the project with 3 next experiments
Each `EXP-NEXT-###` must have the CMP that revealed it, the one thing to change, primary and side-effect metrics, success/failure criteria, the smallest build scope and why it's prioritized.
"Improve the design" isn't an experiment. Make one change and one main metric at a time.
PASS GATE
Mission 4 done check
0 / 8
EVALUATION · 100 POINTS
We grade the links between evidence, more than results