← Back
Generate

Click a step to jump to it

Workfx.AI · Content Creation · Shipped in beta, Aug 2026

Workfx.AI — steering AI content before it’s written

Users were fixing AI drafts after the fact because the choices that shaped them were hidden. I redesigned Content Creation around three decisions users make before anything is generated.

RoleUI/UX Design
Information Architecture
Interaction Design
Vibe coding
TimelineMar – Aug 2026
5 months
TeamSarah Yang · PM
Yueqi Men · PM
Michelle Chen · Designer
PlatformDesktop web
SaaS · SEO / GEO
01

Overview

Moved the decisions an AI made for every draft to before generation, where users can make them.

Workfx.AI is an AI-powered content growth platform that helps businesses find search opportunities and create marketing content for SEO and GEO (Generative Engine Optimization: getting content cited in AI answers such as ChatGPT and Perplexity).

My contribution

Design

Owned the information architecture and interaction design of the three-step flow: which decisions to surface, how to group them, and what to call them.

Designed how users review and manage multiple drafts at once, after my PM scoped batch generation.

Prototype

Built three working prototypes with Claude Code, so design reviews reacted to real interactions, not mockups.

Team

Worked with PMs Sarah Yang and Yueqi Men, who defined what the assistant needed to do, and designer Michelle Chen.

Plus,

Also designed Workfx.AI’s official website and homepage.

Skip to solution ↓See the results ↓
Problem38%

of drafts changed direction after generation, because the choices that shaped them were hidden.

Change5 → 3

hidden system stages became three decisions users make before anything is written.

Result75% · 82%

chose content unaided · revised their choices before generating, in an 85-user test.

02

Problem

Found that everything shaping a draft happened out of sight, so users fixed it afterward.

When I joined, Content Strategy already recommended topics and explained why they were relevant. Users chose a topic, platform and language, generated a draft, and refined it through AI chat.

Every generation costs credits and time. When a draft misses the angle, users don’t just lose a draft: they start to treat the AI as a slot machine. For a platform that sells content growth, that is the moment trust drops.

Content Strategy when I joined: pick a topic, generate, fix in chat.
03

Research

Traced late fixes back to choices users never saw being made.

The project started from beta user feedback on Content Strategy, shared with me by my PM.

Insight 01

Defaults were accepted, not chosen

I just went with the suggestion. I couldn’t really tell what changing those settings would do to the article.
Design response

Each suggested direction says what it changes: the audience, their pain point, and the search opportunity behind it.

→ 01 Choose a direction
Insight 02

Direction was fixed too late

Once I saw the draft, I realized it wasn’t the angle I wanted. So I went back into the chat to explain it.
Design response

Direction, prompts and titles are all chosen before generating, and stay editable until then.

→ 02 Select prompts
Insight 03

System terms weren’t user terms

I know who I’m writing for and what I want to say. I’m less sure what you want me to choose under ‘intent.’
Design response

Every step leads with the user’s action: “Choose a direction”, not “Intent Group”.

→ 01 Choose a direction

Paraphrased from beta user feedback, shared by PM.

What this looked like for one user

Meet Mia, a marketing lead at a B2B SaaS company, writing about AI search visibility. She picks a suggested topic, hits generate, and reads a draft aimed at beginners. She wanted a comparison for buyers. Three chat messages and 4 regenerations later, she has something usable.

Generate → Read → Fix in chat ↻

Drafts reworked in chat38%direction changed after generation
Regenerations per draft4×before users kept one
The loop

Generate → Read → Fix in chat ↻

Where the draft actually got decided

I mapped what users saw against what the system decided. Four of the five decisions happened out of sight.

What users didWhat the system decided behind it
Select a topic01Audience
Choose settings02Intent03Keywords
Generate04Query
Refine through AI chat05ContentThe only stage users saw

My working hypothesis

Users weren’t choosing the defaults. They were accepting them, because they couldn’t see how those choices shaped the result. So control happened after generation instead of before it.

AudiencedefaultIntentdefaultKeywordsdefault ? Generate Change the angle to… Fixed after generation

How other tools handle setup

I looked at how Jasper, Copy.ai and Notion AI handle setup. Most either hide it entirely or turn it into a long form. Neither matched what our users asked for.

Notion AI
Notion AIOne blank prompt. Audience and angle exist only if users think to type them.
Copy.ai
Copy.aiSetup lives as [brackets] inside a prompt template users have to edit.
Jasper
JasperVoice, topic, length and outline: every setting asked before the first word.

Screens via Mobbin.

04

Challenge

Owned how the assistant works for users, while my PM defined what it does.

Which decisions to surface, how to group them, and what to keep in the background. Two tensions shaped every decision that followed.

The tension

Transparency vs clarity

Show every pipeline stage, and internal logic becomes the user’s job.

The trade-off

Control vs speed

Workspace already gave users a fast, chat-first way to get a draft. That freed Content Creation to slow down where it mattered and ask for decisions before generation.

HMWHow might we give users meaningful control over what the AI creates, without asking them to understand the system layer by layer?
05

Design

Turned five system stages into three decisions users make before generation.

The shipped flow, end to end.

01Choose a direction

Each suggested direction explains why it exists (audience and pain point) and shows the search opportunity behind it: related keywords, total searches and competition. If none fit, users add their own.

WhyBeta users said they didn’t know what to pick under “intent”. Explaining each direction in their terms lets them choose it, not just accept it.

02Select prompts

Picked prompts stay pinned at the top, each tagged with its funnel stage and intent group. Below them is a pool of real user questions to add from.

WhyIn V2, earlier choices scrolled away in the chat. Pinning them keeps the current selection visible while users explore.

03Shape the output

Users filter the title pool by content type (comparison, scenario, risk alert, deep dive), add their own title, then set platform, language and how many articles before one batch generate.

WhySEO/GEO content usually targets one question, so each prompt becomes its own title, and one batch action keeps a multi-draft run quick.
Generation progress
04 Each draft reports its own progress, so the wait stays tied to the choices users just made.
AI search tracking
05 Users can opt in to tracking how the content shows up in AI search, then find drafts in a content calendar.

Back to Mia

She chooses the direction “Buyers comparing vendors”, picks the 3 real questions they search for, and keeps one title per question. 3 drafts generate together, each already pointed the right way.

Direction → Prompts → Titles → Generate

How it fits with Workspace

Content Creation doesn’t replace Workspace. They serve different needs.
WorkspaceTask-orientedContent CreationDecision-oriented
01Starts with a conversational promptStarts with selecting a content direction
02Quickly initiates a creation taskGradually defines personalized content
03Presents results through conversations, tasks, and assetsOrganizes the process around intent, questions, and content
06

Process

Prototyped three rounds in code, each answering one question.

V1Which decisions matter?Showed every stageFelt like one question asked three times V2How should they be described?Five stages became threeClearer, but earlier choices scrolled away V3How do users keep track of them?A step-by-step panelShipped · 75% chose unaided

The system was already making five decisions for every draft. What if users made those decisions themselves, before a single word was written?

Sketch: the pipeline as a conversation
The idea that became V1: one question per stage, with a progress track. Early sketch, redrawn.

I weighed three routes against the beta feedback.

Route A

Smarter defaults

  • Nothing new to learn
  • Only improves the guess
  • Users still accept instead of choose
Route B

A settings form

  • Every choice is visible
  • Turns the pipeline into a configuration job
  • Same “what is intent?” confusion
Route C · chosen

A guided conversation

  • Users already chat in Workspace
  • The AI proposes options, so no blank fields
  • One decision at a time

Two rules came with it: decide before you generate, and pick, don’t fill.

Three rounds, one question each

I built all three prototypes with Claude Code, so each version was a working, clickable flow rather than static screens. V1 and V2 were reviewed internally, so early decisions rested on our judgment; the first round with users was the 85-user test on the final flow.

Prototype debug bar
One HTML file, many flows: a debug bar switched between the V1 and V2 paths during review. What vibe coding changed for you, e.g. how fast an idea became clickable.

Iteration 01

V1Showing every stage

The first iteration turned the five-stage pipeline into a guided AI conversation. A progress track (Audience → Intent → Keywords → Query → Content) showed where users were, and at each stage the assistant proposed options.

V1 screen12
  1. 1A five-stage track: Audience → Intent → Keywords → Query → Content.
  2. 2At each stage the assistant proposed options to choose from.

Design review

PM and engineering spotted the problem: three of those stages described the same opportunity from different angles. Users were asked about it three times before seeing any content.

IntentKeywordsQuery→One opportunity
“Intent, keywords and query feel like the same question asked three times.” — PM, V1 design review
Key learning

Making every stage visible didn’t make the decisions clearer.

Iteration 02

V2Grouping inputs into a content direction

V2 kept the conversation but reduced five stages to three. Each direction bundles an intent with its keywords, users pick real search prompts as targets, and each prompt becomes its own titled draft, since SEO/GEO content typically targets one search question.

V1 · five layers AudienceWho are we creating for?IntentWhat do they need?KeywordsWhat topics define the opportunity? Prompt MatchWhat questions should we address?Content GenerationHow should the content be created?
→
V2 · three stages Stage 01Intent Group — choose a content directionAudienceIntentKeywords Stage 02Prompt Match — select questions to address Stage 03Content Generation — define angle & format
BeforeUsers reconstructed context across multiple steps.
AfterUsers understand a content opportunity in one place.
V2 screen12
  1. 1Five stages became three: Intent Group → Prompt Match → Content Generation.
  2. 2Each direction bundles an intent with its keyword group.
Key learning

Grouping related system concepts around a user goal made each decision easier to understand. But in a chat thread, earlier choices scrolled out of view.

Iteration 03

V3From a conversation to a structured workspace

V3 moved the three stages out of the chat into a step-by-step panel. Completed steps collapse, so users always know where they are and can go back to any step. Each step leads with the user’s action, not the system’s term: “Intent Group” became “Choose a direction”.

V3 screen123
  1. 1Steps lead with the user’s action: “Choose a content direction”.
  2. 2Each direction explains itself in plain words: audience, pain point, opportunity.
  3. 3Other steps collapse, so users always know where they are and can go back.

My PM scoped batch generation; I designed how users review and manage multiple drafts at once.

Key learning

Giving users more options was not enough. They also needed a clear way to see and manage what they had already chosen.

07

Impact

Shipped in the Aug 2026 beta, with users making the calls before generation.

Unmoderated usability test with 85 users from Workfx.AI’s partially opened beta, run by the Workfx.AI team on the final three-step flow. Content Creation shipped as part of the beta in Aug 2026.

75%

completed content selection without assistance

82%

correctly identified and revised their choices before generation

Tasks: choose a direction, select prompts, and adjust choices before generating.

08

Reflection

Learned that clarity comes from choosing what to show, not from showing more.

For the design

Transparency isn’t the same as clarity. V1 showed every stage of the pipeline and was harder to use because of it. The work wasn’t exposing the system; it was deciding which parts deserved a user’s attention, in the user’s words.

For the flow

People need to see what they’ve already chosen. V2 made each decision clearer, but earlier choices scrolled away. More options only helped once users could see and manage their current state.

For the process

Prototyping in code changes the review. Working flows let PM and engineering react to real interactions, which is how the V1 repetition surfaced before any user saw it.

What I’d validate next

Testing showed users could make and revise these choices. It didn’t show whether earlier control leads to better content. Next, I’d compare drafts from the new flow against the old one: how often users still rewrite the direction in chat afterward, and whether they keep coming back to Content Creation over Workspace.

The hardest part wasn’t showing users what the AI does. It was choosing which few decisions were worth their attention.

Next On