Nomi · RedNote creator MVP · case study

Redesigning Nomi as a RedNote creator MVP

Nomi tests whether creating through one deliberately shaped social self produces better content than direct prompting. I designed and validated the concept across two research rounds, moved the MVP to RedNote, cut the flow from eight screens to four, and resolved eight design questions from my original wireframes. This case study runs from research through outcome and ends with a working prototype.

Solo designer and researcher, ZeyaAI Round 1 n=16 · Round 2 n=16 July 2026
Research

How high-engagement RedNote notes are written

Before designing generation, I studied the platform's content economy: what the algorithm rewards, how top notes are structured, and why. Three findings shaped the product directly, and I wrote format studies to internalize the craft before asking a model to reproduce it.

Finding 01

Titles carry a hard limit of about 20 characters and lead with value: a number, a mistake avoided, a comparison, a hidden gem. The title and cover together decide the click in a two-column feed, so the hook is a design constraint, not a style choice.

Finding 02

The algorithm scores engagement by weight: a like counts 1, a save 1, a comment 4, a share 4, a follow 8. Saves and comments, not likes, decide whether a note leaves its first traffic pool, and the first hours after posting size that pool.

Finding 03

Save-earning notes share an anatomy: a personal hook, a numbered utility list, a reason to save for later, and a question that invites comments. Roughly a 60/40 split of practical value to personal narrative performs best.

Finding 04

RedNote doubles as a search engine. Keywords belong in the title and first lines, and 5 to 8 tags mix broad and niche terms. This is why the product surfaces tag view counts and keyword-aware titles instead of leaving them to chance.

Format studies

Three notes I wrote to the platform's conventions, one per register the product generates. Each label names the technique the note demonstrates.

cover: badge and morning light at Apple Park
my day as an Apple intern in the Bay
6:40am. Caltrain, cold brew, rehearsing my standup update in my head. what nobody tells you about interning here: 1. the campus is a circle and you will get lost in a circle 2. lunch is subsidized, the cafe matcha is not a secret anymore 3. your badge opens fewer doors than your questions do 4. design reviews are quiet. the silence is the feedback ask me anything about the recruiting timeline, I kept every email. save this for your application season.
#applestudentintern #bayarealife #techinternship #dayinmylife
day-in-the-life hooknumbered listsave CTAcomment invitation
cover: tasting flight on a wooden tray, Uji storefront
5 matcha stores in Japan worth the flight
I planned a whole trip around matcha and regret nothing. 1. Nakamura Tokichi, Uji. the fresh mochi arrives still warm 2. Ippodo, Kyoto. buy the koicha, thank me later 3. Tsujiri Honten, Gion. the parfait is a structural achievement 4. Suzukien, Tokyo. gelato in seven intensities, take level six 5. Saryo Tsujiri. warabi mochi that ends arguments queue times and what to order first are in the last image. save this before your Japan trip disappears into a screenshot folder.
#japantravel #matcha #kyotofood #mochi
number hookpersonally testedreference detailsave CTA
cover: suite window facing the Matterhorn
8 Swiss resorts where the view does the work
I compared 23 resorts so you only have to choose from 8. 1. The Cambrian, Adelboden. the pool photo is real 2. Riffelalp, Zermatt. Matterhorn from the outdoor spa 3. Burgenstock, Lake Lucerne. the funicular is the check-in 4. Villa Honegg. yes, the one from the video 5. The Chedi Andermatt. a cheese tower taller than me 6. The Omnia, Zermatt. cliff elevator, quiet rooms 7. The Alpina Gstaad. quiet money, loud mountains 8. Le Grand Bellevue. a wellness circuit worth a full day rates drop hard in shoulder season, late September is the move. which one is your type? comment your pick.
#switzerland #luxurytravel #alpineresorts #travelguide
curation credibilitycomparison listinsider tipcomment prompt

What this means for Nomi: generation has to target the whole anatomy, hook title, list body, save line, comment prompt, tag mix, and the guidance layer I call the recipe carries the algorithm knowledge, publish timing and tag selection, so the user's self can carry the voice.

Ideation

The original concept and its wireframes

My first wireframes cast Nomi wide: a mode select across three jobs, a profile builder that collected personal facts, a separate context page, and a generation screen with a detail editor. The flow below is the starting point this case study revises; the per-screen references appear in Iterations next to what replaced them.

Original flow, eight screens and three modes
Mode select3 paths
→
Profile picture generation
→
Profile builderpersonal facts
→
Context dropseparate page
→
Generate 3 cards
→
Detail and edit
→
Networking messagesparallel path

Two setup pages ran before any value appeared, three ambitions shared one product, and content extraction arrived after identity was already fixed, where it could not reduce setup cost. Testing made each of these costs measurable.

Testing

Two validation rounds, 32 participants

Both rounds used a within-participant comparison: the same person, the same topic, direct generation against social-self generation, then forced choices on fit, postability, and editing burden, plus save and reuse intent.

Round 1 · any platform

Tested whether a shaped self matters at all. Demand for saving and reusing a self was the strongest signal; platform breadth diluted output quality.

intended-self fit50%
more likely to post56%
open to saving the self81%
would choose the workflow again63%
Round 2 · instagram only

Same comparison, one platform, lighter setup. The three core output metrics improved. Time to a usable result did not move, and save intent dipped.

intended-self fit69%
more likely to post63%
less editing required56%
faster to usable result44%
Insights

What the numbers meant, and the five decisions they produced

The rounds proved the thesis and localized two failures. Setup feels like work before it shows value: the speed metric stayed flat at 44% while everything else improved. And persistence is stated but not felt: 75 to 81 percent said they would save and branch a self, yet nothing in the product made that saved self visible or rewarding. Each decision below attacks one of the two, inside a RedNote-only frame chosen for its creator-identity economy, its US-entry timing, and its mechanical parallels to the Instagram evidence.

D1

RedNote content is the only mode

I cut networking messages and profile appearance. Round 1 showed that breadth across contexts diluted every signal, and only content creation was ever validated.

The self object stays platform-agnostic underneath, so other modes can return in Nyla without rework.
D2

One self, with a visible entry to Nyla

Nomi is a one-self trial. The home screen shows the saved self plus a locked card for a second one that opens a waitlist prompt. Wanting another self becomes a counted event instead of a survey answer.

75% said they wanted another self; this converts stated intent into observable demand.
D3

Three seed sources replace the bio form

A self starts from the user's own posts, from a creator they admire, or from a RedNote archetype, then three dials and one identity line. The original form collected facts the model could not use as voice.

Round 2 improved by lightening setup; seeding from real posts removes prompt writing entirely.
D4

A live voice preview during setup

A sample note re-renders on every adjustment while shaping the self. The user hears the self before generating anything, so setup produces value instead of asking for faith.

Aimed directly at the flat 44% speed metric.
D5

Platform anatomy, user covers, and the recipe

Drafts are structured as real notes: hook title, list body, save line, tags chosen by the research findings. Covers are uploaded, never generated. Nomi's guidance layer, the recipe, appears only at service moments.

Platform-native output drove the 19 point fit gain in Round 2; the recipe carries the algorithm knowledge.
Iterations

From eight screens to four, with each revision argued

Revised flow
01 Shelfthe saved self
→
02 Shapefirst run only, under 90s
→
03 Composeone input
→
04 Studiodrafts, triage, copy out
Returning users skip 02. Accepted edits in 04 teach the self. The locked card on 01 counts Nyla demand.

Three input decisions I standardized across builds

Across versions of Nomi, the same three input questions kept getting re-answered: by me, and by the AI pair tools I built with. The interim build below shows both patterns mid-evolution, redrawn in light mode. I turned the answers into rules and applied them everywhere.

Interim build: profile as a form, pills arriving
Your creative profile
This is how Nomi learns to sound like you. Fill in as much as you like.
Gen AIDigital AssetsCryptoPredictive AIConversationadd topic
INFJ Gemini. I am an introvert who enjoys me time. I write diaries and love to collect shells.
Who I like
Paul GrahamAndrej KarpathyLi Jin
Platforms I use
RedNoteLinkedInInstagram
Save profileContinue
Interim build: five variations, prompt on the same screen
Variations
Marginal Benefits of AI in ProdAI product design delivers significant marginal benefits by accelerating iterative processes.
Big Data Modeling in PredictivAI product design delivers significant marginal benefits by accelerating iterative processes.
Big Data Modeling in Human InAI product design delivers significant marginal benefits by accelerating iterative processes.
AI Automation in Design CyclesHow AI automation is streamlining the iterative cycles of product design.
The Front-End Reliance of AIA deep look at why AI concepts depend so heavily on front-end human interactions.
Predictive AIAI Visualization I want to create a series of notes that investigate the marginal benefits of AI in product design. Proceed
Rule 1

Where prompt pills belong

I add pills where intent is known but phrasing is not: topics, reference creators, warm starts, refine scopes, trait words. Recognition replaces recall, which attacks the prompt-skill prerequisite from my research notes and the flat 44 percent speed metric, and every tap lands in analytics as a countable choice where free text cannot.

The one input pills never replace is the note idea itself. The topic is the single thing the system cannot know for the user.
Rule 2

Why five variations became three

In the interim build, three of five variations opened with the same sentence, so choosing meant proofreading. Round 1 had already shown that people respond to versions of a self, not shuffles of one text. Three drafts in three registers make the choice about identity, fit on screen without scrolling, and turn each keep or discard into a clean per-register preference signal.

Fewer, more distinct options is also what the forced-choice study design itself relied on.
Rule 3

Where input lives, and where it is removed

Input sits where its consequence is visible, and nowhere else. The note idea sits beside generation, refine sits under the drafts it rewrites, identity controls sit beside the live voice check. Where the consequence is invisible the input is removed, which is why profile editing left the compose path and two setup pages became one box.

Every page before value is attrition surface in my drop-off metric. Merging input with its preview turns prompting into steering.

01 Shelf: the home screen holds the saved self

replaces the mode select home and the accounts list
Original: mode select home
nomi
RedNote Content
Networking Message
Profile Appearance
Continue
What I changed
  • Removed the mode choice. The validated job, RedNote content, is the whole product.
  • Replaced the empty landing with the saved self, so the product's memory is the first thing a returning user sees.
  • Added a locked second-self card that opens a Nyla waitlist prompt, converting branch intent into counted clicks.
Choose a self to write with
A self is your saved RedNote voice. It remembers how you sound, so every note starts from you.
Rain
"Uji at first light, warm mochi in hand. everything since has been an audition."
12 notes in, drafts start closer to done.
Every note you keep teaches Rain more of you.
Write with Rain
locked
New self
Nomi is a one-self trial.
Nyla holds as many selves as you need.
join the Nyla waitlist
Recent notes
5 matcha stores in Japan worth the flightcopied to RedNote · 2d ago
a slow morning in Uji before the crowdskept as draft · 5d ago
Reshape Rain
1 · Persistence first

The saved self is the home screen

Saving and reusing was the strongest signal in both rounds, so I made it the architecture rather than a feature. The card leads with a line in the self's voice because people recognize a voice faster than they read metadata.

2 · The Nyla entry

A locked card that states the limit honestly

One self in Nomi, by design. The card explains the limit and links to the Nyla waitlist. Each click is a measured unit of multi-self demand, and doubles as a willingness signal for the business model.

3 · Publishing stays manual

Recent notes record copy events, not posts

The MVP posts nothing. A note's life ends here as copied to RedNote, which is a proxy I can count honestly without account integration.

02 Shape: seeding, dials, and a live preview on one screen

replaces the profile builder and the separate extraction step
Original: profile builder
Rain
avatar
Intro: INFJ Gemini. I am an introvert who enjoys me time. I write diaries and love to collect shells. My favorite ice cream flavors are blueberries, hazelnut.
Gen AIDigital AssetsCryptoConversation
What I wish to be:
Who I like:
Social platforms I use:
Save this profileContinue
What I changed
  • Replaced personal facts with voice controls. The facts described a person; they did not control how output sounds.
  • Moved extraction upstream: a self now starts from the user's own posts, a creator they admire, or an archetype.
  • Added a live sample note that re-renders as the dials move, so nobody sets up a voice blind.
  • Held everything to one screen with a 90 second budget. The original spread setup across two pages.
step 1 of 1 · about 90 seconds
Seed
Where I am
The voice you have today. RedNote posts work best, but any writing counts: captions, blogs, notes, screenshots.
Where I want to be
Inspirations: upload anything that pulls you forward. Nomi borrows the register, never the identity.
Categories, in both
foodtravelhonest reviewsslow living
Tune
quiethigh energy
useful for readersdiary close
off the cuffstructured
Your own words: warm, picky, detail-driven. Add or remove traits.
Anything else that defines you: whatever the options above did not cover.
Hear it
voice check · an example, not your content
drop everything, new bakery
this bakery broke me a little warm bread at 8am and the owner remembered my order this is my whole personality now
#bakery #morningfind
a sample topic, so the voice is judged before it touches your notesSave Rain
1 · Two directions

Where I am, and where I want to be

The seed resolved into an identity question rather than a data source. Where I am holds the voice the user already has, from any platform, pasted or as screenshots. Where I want to be holds inspirations: anything that pulls them forward, with the register borrowed and the identity left alone. The old archetypes dissolved into category options inside both, so aspiration is expressed by category, not by template.

2 · Sliders plus words

Spectrum controls for the dimensions, the user's words for the language

I tried replacing the sliders with binary word pairs and reversed it: pairs collapsed the spectrum users actually feel. The final control is both layers. Sliders carry the three dimensions, and a trait field carries the user's own words, which testing showed is how people describe a version of themselves. One identity line remains as the ownership ritual.

3 · The voice check

The preview re-voices labeled examples, never your content

The live sample is an example note on a neutral topic, labeled as such, with more examples one tap away. Judging the voice on neutral content isolates the one variable this screen controls, keeps shaping from sliding into composing, and still removes the prompt-skill prerequisite: users steer against visible output. Because the example is not theirs, the save action carries the weight, so it anchors the whole screen.

03 Compose: context and intent merge into one input

replaces the context page and the separate create page
Original: context page, then a create page
Rain · INFJ Gemini. I am an introvert…Update Looks
Give me some context of what you are creating today:
Drop here
Give me anything. Input a link to your most recent work, a paper you wrote, a conversation, books you have read, transcripts from conferences.
Accounts I manageContinue
What I changed
  • Merged the context drop and the creation prompt into one input. To a creator they are the same sentence.
  • Reduced the self to a switchable chip. Identity editing does not belong on the path to writing a note.
  • Removed the accounts list; multi-account management belongs to Nyla, not this trial.
  • Added warm starts: a series continuation, one trending pull, an unfinished draft.
Rainwarm · high energy · honesttap to reshape
What is the note about today?
Type it, paste it, or drop it: an idea, a link, a photo set, a transcript. Rain reads anything. Output is English by default.
add fileadd link
Generate 3 notes
continue: Japan matcha series, part 2recipe: trending in your niche, Kyoto autumnresume: saved draft from Tuesday
1 · Self as context

The chip introduces the self and opens Shape

Returning users reach this screen in two clicks. The chip carries the self's name and its traits, so anyone can tell at a glance who is writing, and tapping it goes straight to Shape for reshaping. The Nyla door stays on the shelf, where the demand count lives.

2 · One input

Fewer pages before value

My drop-off metric made the argument: every page before generation is attrition surface. Per-note context stays per note; only what the user explicitly teaches persists to the self.

3 · Warm starts

The recipe suggests, the user decides

These are labeled as recipe suggestions and measured by tap-through. Nomi's guidance layer serves; it does not narrate.

04 Studio: drafts in feed context, edits that teach the self

replaces the generation cards and the separate detail and edit screen
Original: generation cards, then a detail screen
What changes would you like to make? · Texts / Image toggle
check mark · card stack · cross mark
#PredictiveAI · 1M views
#GenerativeAI · 3.1M views
#BigData · 900K views
This post performs best at noon on workdays.
What I changed
  • Merged generation and editing into one screen and promoted the feed preview to a fixed panel.
  • Made the three drafts three distinct registers of the note rather than three variations of one text.
  • Replaced the swipe stack with keep and discard buttons; the job was triage and buttons do it with all drafts visible.
  • Replaced the image toggle with a cover upload; the product does not generate images.
  • Routed all learning through a visible teach control; nothing trains silently.
Rain · 3 drafts · English
Your RedNote page preview · drafts box
your RedNote page, simulated from a screenshot
5 matcha stores in Japan worth the flight
upload cover · concept: tasting flight, Uji storefront
5 matcha stores in Japan worth the flight
I planned a whole trip around matcha and regret nothing. queue times and what to order first below.
#japantravel 2.1M · #matcha 1.4M
keepdiscard
upload cover · concept: twelve desserts, scored flat lay
I tried 12 matcha desserts in Kyoto
criteria: grind, sweetness honesty, texture, price. number 4 will start fights. scores below.
#kyotofood 860K · #matchareview 640K
keepdiscard
upload cover · concept: empty street, first light, one cup
a slow morning in Uji before the crowds
no list today. one street before the shops opened, one warm mochi, one small routine.
#slowtravel 920K · #diary 3.4M
keepdiscard
Refine: toneword choiceattitudeshorter teach Rain from my edits
recipe: notes like this perform best Thursday to Sunday, 12 to 2pm PT
Download Copy to clipboard
1 · Page preview

The panel simulates your own page, not a generic feed

Seeing the note inside your page is the sharpest version of the question my study asked: is this me, here? Users can upload a screenshot of their current RedNote page to simulate the full grid, and a drafts box holds every unposted note as cards, so several can be placed, dragged into order, and previewed together, with the current one highlighted. Cells follow RedNote's real card anatomy: 3:4 cover, two-line title, author row.

2 · Note anatomy

Three registers, covers by upload

The drafts are a guide, a review, and a diary entry, which forces a real choice. Covers are user photographs, never generated; the concept line describes the shot until one is uploaded.

3 · Triage and refinement

Keep, discard, and three scoped rewrites

The tone, word choice, and attitude scopes come from my original board notes. Keep and discard preserve the triage intent of my swipe stack with every draft visible and each click legible in analytics.

4 · Visible learning

Edits teach the self only when the user says so

The teach control is the second learning moment after seeding. It converts editing labor into compounding value, and it keeps consent explicit: triage clicks are choices, not training data.

Eight design questions I raised and resolved

Reviewing the original wireframes against the research produced eight open questions. I resolved each with a decision and priced it as a trade-off.

01 · scope

Whether networking and appearance modes belong in the MVP

QuestionMy wireframes carried three parallel modes, but only content creation was ever validated by the research.

DecisionBoth extra modes are out of the MVP. The self object stays platform-agnostic underneath so a networking self can exist in Nyla without rework.

Trade-offI lose the chance to discover networking demand accidentally in round 3. I accept that, because accidental discovery on unvalidated surfaces is what muddied Round 1.

02 · the bio form

What the personal-facts form was reaching for

QuestionThe original profile builder collected MBTI, favorite flavors, and interests. Reviewing it, I recognized the real intent: giving users a blogger to look to as a starting point for becoming one.

DecisionThat intent became a first-class feature. A creator you admire is one of three seed sources. The system extracts the reference's register only, meaning rhythm, structure, and energy, never name or story.

Trade-offAspiration seeding solves the cold start for users with no posts, but risks sound-alike voices. Guardrails: register-only extraction, a mandatory identity line, and the reference stays private to the user.

03 · covers

How a cover-led platform works in a text-first MVP

QuestionRedNote discovery is driven by covers, but both research rounds deliberately excluded image generation so it would not confound the voice comparison.

DecisionUsers upload their own cover to preview it in the feed; otherwise the draft shows a written cover concept with an upload control. The product does not generate images.

Trade-offUploading is a manual step and depends on the user having a photo. I track the upload rate: if it stays low, the feed preview loses power, and that is the signal to revisit generation after the voice loop is proven.

04 · language

Default language for US-entry creators

QuestionThe product faces US-entry creators, so I had to choose between English, Chinese, and a bilingual toggle.

DecisionEnglish is the only body language in the MVP. Tags remain mixed English and Chinese because that is how discovery works on the US feed and it costs the user nothing.

Trade-offA simpler product and a cleaner round 3 cohort, at the price of excluding bilingual creators. If recruiting surfaces them in numbers, the toggle returns; the generation layer already supports it.

05 · one self

How many selves the trial allows

QuestionMultiple selves are the Nyla vision, but an MVP that ships them would be testing management before proving the value of one.

DecisionNomi allows one self. The home screen keeps a visible locked card for a second; opening it states the limit and links to the Nyla waitlist. The compose screen repeats the same entry in passing.

Trade-offRound 3 never observes real multi-self behavior, which Nyla ultimately needs proven. In exchange, every waitlist click is a counted, attributable demand signal, and the MVP stays honest about what it is. A smaller true measurement beats a larger fake one.

06 · triage

What the accept and reject stack was for

QuestionMy original detail screen had a card stack with accept and reject marks. The intent was draft triage: letting users select and make choices.

DecisionKeep and discard buttons on every draft. Same choice, three practical gains: all drafts stay visible during the choice, the controls are keyboard accessible, and each click lands in analytics as a per-register preference.

Trade-offButtons are less tactile than swiping. On a future mobile build, swipe returns as a gesture over the same keep and discard model; the data schema does not change.

07 · monetization

Where the subscription signal lives in a free study

QuestionMy metrics notes listed subscription conversion, but a paywall inside a 16-person study loop would contaminate every completion metric.

DecisionThe loop stays free. Willingness is measured twice, never as a gate: the Nyla waitlist clicks, and one post-study question asked at a concrete moment, whether the user would pay to keep Rain.

Trade-offStated willingness is weaker evidence than a real conversion. I take the honest weaker signal now; pricing gets its own test when Nyla exists.

08 · the recipe

Whose voice the product itself speaks in

QuestionMy early wireframes had the assistant commenting between steps. The concept underneath: users define explicit selves and own them; Nomi has no explicit self of its own, but it has an implicit one, the native recipe that boosts engagement. The selling points are management and the perks of that recipe.

DecisionThe recipe became the product term. It appears exactly where it serves: warm starts on Compose, publish timing and tag rationale in Studio, always labeled and visually distinct, never as conversation between the user and their draft.

Trade-offAn ambient guidance layer demos less flashily than a talking assistant. But a narrator would make the tool the author, and the evidence says the value is the user's own self, elevated. The recipe earns visibility through measured tap-through, not personality.

Prototype

Interactive prototype of the four screens

The revised flow, working end to end. The self is editable: rename it, choose a look for its orb with live preview before confirming, seed it from where you are and where you want to be, and hear labeled example notes re-voiced as you tune, with saving as the screen's one commitment. The studio simulates your RedNote page, from a screenshot if you upload one, holds every unposted note in a drafts box styled like the page itself, lets you drag notes into any order, and ends in copy to clipboard with download beside it. Use the arrows or the step pills in the top bar to move between screens. Generation is canned to one sample idea, a Japan matcha and mochi trip, so the step is believable without a model behind it.

nomi · rednote creator

Choose a self to write with

A self is your saved RedNote voice. It remembers how you sound, so every note starts from you. Pick one to write, or reshape it first.

Rain
"Uji at first light, warm mochi in hand. everything since has been an audition."
12 notes in, drafts start closer to done.
Every note you keep teaches Rain more of you.
locked
New self
Nomi is a one-self trial.
Nyla holds as many selves as you need.
Recent notes

Shape Rain

reshape any time · about 90 seconds
Seed
The voice you have today. RedNote posts work best, but anything you have written counts: captions from any platform, blog posts, messages, notes.
What you share today
Tune
quiethigh energy
useful for readersdiary close
off the cuffstructured
traits the sliders cannot say, folded into the voice
suggestions, tap to fill the box
your own words, fed straight to the voice
Hear it
voice check · an example, not your content
Example topic: a bakery find
drop everything, new bakery
this bakery broke me a little warm bread at 8am and the owner remembered my order this is my whole personality now
#bakery #morningfind
a sample topic, so you can judge the voice before it touches your notes
Nothing changes until you save. Saving updates how Rain writes everywhere.

New note

What is the note about today?

Rain · 3 drafts

three registers of the same idea · English
Your RedNote page preview
your RedNote page, with this note placed in it
Refine:
teach Rain from my edits
recipe notes like this perform best Thursday to Sunday, 12 to 2pm PT
writing as Rain
Copied

Prototype boundaries: generation is canned to the sample idea, the voice check re-voices labeled example topics and never user content, refine controls swap written variants, custom refines are acknowledged but not applied, uploads never leave the page, the avatar recolors are placeholders until the final five colors arrive, and the waitlist link counts a click and goes nowhere. Navigation, renaming, the look preview and confirm, triage, the drafts box with drag ordering, the teach control, copy, and download are real. A standalone copy of this prototype lives in its own file for testing sessions.

Outcome

Instrumentation for round 3 and the path to Nyla

The redesign converts every open question into something round 3 can count. The funnel below maps each metric from my research notes onto a mechanism in the revised flow; the two decisive numbers are return-with-the-same-self, which tests whether persistence became real, and Nyla entry clicks, which test whether one self creates demand for many.

Self created
time to save and seed source chosen; covers the extraction-completion metric and the unresolved setup-speed gap
First generation
completion rate and keep or discard per register; covers generation completion and drop-off
Cover uploaded
upload rate per kept draft; decides whether cover generation ever earns a place
Note copied out
copy events per session, the honest proxy for posting
Return with the same self
second session within 7 days; the willingness-to-return metric that decides Nyla
Self taught
teach acceptances per session; makes persistence measurable rather than stated
Nyla entry clicked
locked-card opens and waitlist joins; branch intent becomes counted demand and a willingness signal

If round 3 shows returning sessions and waitlist demand, the next build is Nyla: the same self object, multiplied and managed. If it does not, the one-self loop still stands as a validated creator tool, and the research record shows exactly where the thesis stopped.