Product design / AI workflowOngoing · 2026

Finfold

Finfold fish mark

One update
seven rewrites

X wanted a hook. Xiaohongshu wanted the everyday problem first. LinkedIn wanted the backstory. I rewrote one product update seven times before I could post it everywhere. Finfold started there: AI drafts for each channel, and a person makes the final call.

AI product design · Interaction · Interface systemExplore the interface ↓
My role
AI Product Designer · UX & UI design lead
Collaboration
Product and engineering collaboration
Status
Live product · Ongoing iteration
Inaugural Shanghai “AI + Consumption” Innovation CompetitionExcellence Award
SoloVerse3rd place · Emergence Award
FC-OPC × Next iBotBest Edge–Cloud Collaboration
X-Agent AI MCP4th · Open Innovation
↗
01

Overview

One product update
seven rewrites

After shipping a product update, I wanted to post it on X, Xiaohongshu, WeChat, LinkedIn, Reddit and Product Hunt. Copying took seconds. Adapting took forever. A chatbot could draft, but every channel meant explaining the product, the audience and the platform’s unwritten rules all over again. One launch turned into seven rounds of prompt roulette, and the time I saved went straight into checking.

Each time I rewrote it, I noted what that pass was actually fixing. Two kinds of work kept showing up. Names, prices, dates and product limits could not change by a single word. Openings, structure, tone and image sizes had to change for every channel. Everything I designed in Finfold afterwards deals with how those two kinds of content live together.

Product adoption

Already in use

September 2026 · Product data
2,200Cumulative users
50%+Users outside ChinaSee content performance ↗
One source → Channel drafts → Human reviewClick to enlarge

My contribution

I mapped the core tasks, rebuilt the workspace layout, and designed the bilingual light and dark interface system. With the team, I carried the states nobody draws in a demo, like confirm, cancel and recover, all the way into the product.

02

Market research

Writing got faster
results didn’t follow

First I needed to know whether this was just my problem. A 2026 report from CMI and MarketingProfs put two numbers side by side. Most marketers using AI to write felt more productive. Fewer than four in ten felt their content performed better.

87%

Reported improved productivity

39%

Reported improved content performance

B2B marketers using AI for content creation · CMI / MarketingProfs · 2026 ↗

The gap between 87 and 39 is the work I wanted to do

Research scope and methodology

The report covers 1,015 B2B marketers, primarily in North America, surveyed June 24–August 14, 2025. Percentages are self-reported by the subgroup using AI to create content.

03

Competitive analysis

One tool keeps the voice
another keeps the calendar

The best tools in the market each guard one end. Jasper keeps a brand sounding like itself. Buffer puts every channel on one calendar. The stretch in between, from one update to several drafts you can actually post, was still the user’s job to carry by hand.

Jasper ↗

Brand expression

Keep a consistent voice

Reusable context reduces repeated briefing

Buffer ↗

Content operations

Bring every channel into one calendar

Visible schedules and states make handoffs clear

Finfold
Design opportunity

Carry the stretch in between

Brand context → Channel drafts → Human review

↗
04

User research

Two real businesses
one publishing problem

My seven rewrites were a sample of one. In research I interviewed participants and followed two real clients: a bike retailer in Australia and a cross-border medical service. Very different businesses. Almost the same checkpoints between an idea and a published post.

Four parts of a publishing task

Task analysis
  1. 01

    Gather

    Product context and intent

  2. 02

    Adapt

    Opening, length, structure

  3. 03

    Review

    Facts, tone, requirements

  4. 04

    Reuse

    Drafts, brand, outcomes

05

Insights & opportunities

Everyone re-explained their brand
no one wanted to post blind

The same lines kept coming up in interviews. Every session started with re-introducing the brand. Every channel needed its own pass. And at the last step, nobody wanted a machine to press publish for them. Those three remarks became three design principles.

01 / “I start from scratch every time”

Introduce the brand once

Brand memory and rules come along to every session

02 / “Every channel needs its own pass”

Make channel differences visible

Switch channels and edit each draft on its own

03 / “I want to press the last button”

Keep publish in human hands

Review first, then hand off, with consequences visible

06

Architecture & prototyping

Six scattered chores
three connected steps

Broken down, one launch was six separate chores, bouncing between a chat window, a doc and a spreadsheet. The new structure has three steps, with brand memory running underneath all of them. The layout changed too. The old version split source, channels and results into three columns, leaving the draft, the thing you actually have to read word by word, a narrow strip.

Earlier workflow

Six chores, constant switching

  1. 01Find the angle
  2. 02Explain the brand
  3. 03Write the copy
  4. 04Prepare visuals
  5. 05Revise drafts
  6. 06Record feedback
The Finfold workflow

Three steps, the last one is yours

  1. 01

    Bring the source

    Product update · Brand context

  2. 02

    Create and edit

    Content shaped for each channel

  3. 03

    Review and follow through

    Human review · Saved content

Brand memory carries throughPositioning / Audience / Voice / Examples

Earlier layout01

Nav
Source
Channels
Outputs

Three equal columns squeezed the draft into a strip

Revised structure02

Nav
Source + channels
Content preview

Inputs share one side; the draft gets the room

07

UI & design system

Keep the eye
on the draft

Mist and graphite themes share one hierarchy. Glass effects stay in navigation and controls. The writing area stays calm, because this is where people check every word, and any extra shine pulls attention away.

Visual language

Light and dark
with the same hierarchy

Less border weight and fewer competing accents. Space, type, and a restrained green distinguish priorities.

Aa 01Reading first
Clear numbers, quiet hierarchy
08

Design decisions

What AI can carry
and the step that stays human

Getting AI to draft is no longer the hard part. The hard part is deciding what to hand over and which step must stay with a person. This chapter is a handful of those calls, including two things I deleted myself.

01 / Context design

Make brand context visible and editable

Early on, the brand check ran last. All eleven drafts were generated before the system warned that the tone might be off. The warning was right, it just came too late: people had already edited, and starting over felt like punishment. So I moved the rules before generation. How the brand is spelled, which numbers can’t change, which words are off-limits all go into context first, then get checked again afterwards.

Saved voice principle
Clear, credible, without overpromising
Joey Zhao · Person + brand

The design decision

Context you repeat goes into memory; each task’s source stays on the workbench. Brand information lasts for months while sources change every time, so they live in separate places, and people can always see what the AI was working from.

02 / Channel adaptation

One source, two different openings

Two drafts from the same saved content kit, for X and LinkedIn.

Source excerpt

Finfold helps solo founders and small teams turn one product update into platform-specific content drafts.

One product updateTwo channelsReview before publishing
𝕏X / TwitterDraft excerpt
Solo founders don’t hate writing. They hate writing the same update four different ways.
01

Open directly with the repeated work

02

Number the workflow and its limits

Short hook · Numbered thread
inLinkedInDraft excerpt
Most solo founders publish one platform and quietly abandon the rest.

Not from laziness. From arithmetic.
01

Start with the founder’s situation

02

Explain the workflow and design choices

Situational hook · Narrative structure

The structure changes by channel; the facts and review controls stay in view

03 / Honest states

I deleted two things
that made the demo look better

Neither change shows up in a screenshot, and the demo got slightly worse. I made them anyway. People spend years building an account, and one post with a wrong fact lands on them, not on us.

01

A row of pretty fake numbers

The first overview screen showed 42 posts queued, 11 in review, 18 scheduled. It made the product look mature. The numbers were demo data, so the whole row went. An honest empty state is better than a borrowed full one.

02

A fallback that always succeeded

When the model failed, the system quietly slipped in a template and still showed success. Demos ran smoothly, and users had no way to tell which drafts the model wrote and which were filler. Now a failed card says why, can be retried on its own, and every successful result stays put.

Make the starting point obvious

The Agent opens with a prompt and common tasks. Supporting panels appear when needed, so the first screen stays approachable.

Keep frequent actions next to the result

Editing and copying stay within reach. Secondary controls step back so the content remains the focus.

01

Honest progress

Show elapsed time and returned outputs; let people read what is ready

02

Recovery

Keep cancellation, retry, and draft recovery within the workflow

03

Human review

Distinguish drafts, review, and publishing handoff so consequences are clear

04

Six channels at a time

Running all fourteen channels at once looks impressive, but waiting and failures rise together. Six is closer to a real launch batch, and still something a person can review properly.

09

Usability testing

Three tasks
aimed at where things break

Testing skipped the smooth path on purpose. Two of the three tasks were about what happens after something goes wrong: fixing a flawed draft, and leaving halfway then coming back. That is where AI products most often lose trust.

TASK 01

Turn one update
into two channel drafts

Checked first actions and channel selection

TASK 02

Find a problem
and revise one draft

Checked editing controls and version understanding

TASK 03

Return after an interruption
and finish the task

Checked draft retrieval and recovery

10

Outcomes

A post that went out
and four honors

Once the product was live, Finfold content started going out for real. This Xiaohongshu post comes from an actual account in the project records. I show only the single post’s numbers and don’t credit them to the redesign.

Xiaohongshu · Design education

One post
3,078 likes

Why do McDonald’s seats make you want to leave after 40 minutes?

Single-post metrics · September 2026 pitch deck

Project recognition

2026 / FINFOLD

Excellence Award

Inaugural Shanghai “AI + Consumption”
Innovation Competition

3rd place · Emergence Award

HuaQing PuZhi · SoloVerse

Best Edge–Cloud Collaboration

FC-OPC × Next iBot
Shanghai Creative Hackathon

From project presentations and live demos

The habit this project
broke for me

AI products are easy to design for the demo. The happy path is the easiest thing to draw and the easiest way to fool yourself. This time I spent most of my hours on the moments a demo never shows: the model is slow, a draft fails, someone leaves and comes back. None of that makes it on stage, but it’s exactly where people decide whether to trust the product with their account.

The team

The people behind Finfold

Blue team portrait · Design & Product

Design & Product

Product direction and experience design

Blue team portrait · Engineering

Engineering

Core systems and implementation

Blue team portrait · Business & Growth

Business & Growth

Commercial strategy and growth

Blue team portrait · AI Research

AI Research

AI research and applications

Joey Zhao / AI Product Designer

Let’s work together

Open to AI product design roles and project collaborations