Choose an approved template. Create in conversation. Review in context.
Design Center
QA Review
Generate & refine
Approved templates
One foundation. Governed paths from approved content to editable output.
A governed platform that turns approved content into brand-compliant, editable and localized output, without putting a designer in every routine production loop.
Role
Product & UX lead
Period
2024 — Present
Surface
Internal web platform
Scale
Org-wide · CNGS
Status
Mixed · MUSE live, platform in validation
130certified PPT layout modules
4product workflows on one platform
3 daysmanual exhibition deck placement today
System architecture
One entry, four governed workflows, one shared constraint layer.
Future goal · 2027+ Creative Assets Hub — one asset center for outputs, templates and rules, reused across products
The problem
Amazon Global Selling China produces seller-facing decks, campaigns and event materials continuously. Everything routed through a small central design team, so the queue set the pace instead of the idea.
Making slides was never the hard part. The problem was how approved content becomes compliant, editable and localized output without the brand drifting one request at a time.
Product model · MCP
Design MCP is a Model Context Protocol server that holds the design system as executable capability — brand rules, certified templates, module libraries, rendering and QA — and exposes it to any AI client. PPT Studio, MUSE for HTML email, Adaptive Creative and the shared Image Library are four product paths that call the same server.
Why MCP rather than an API or a shared Skill:
APIWrap once per product
Every product has to integrate and adapt the design system separately. Four products, four integrations, four places rules drift.
SkillCopies that go stale
A shared skill is a snapshot of "how we do it", scattered across agents. Once shared, nobody owns the copy; it expires quietly.
MCPOne live source, any client
The latest brand rules, templates, render and QA live in one model-agnostic mechanism. Skills only guide when to call it. Every AI client reuses the same design system, always current.
I own the product definition, interface, template systems, validation rules and the operating model that connects design, engineering, vendors and business teams.
Why it holds up
Brand rules are versioned centrally, so one update propagates to the organisation instead of being re-explained for every request. Each workflow inherits the same logic for approval, QA and human intervention.
My role moved with the system: from taking design requests to owning the standards the system enforces.
The seller conference creates three different deck problems. Approved Chinese sessions need full template application. Global English keynotes need format conversion and image sourcing, while translation stays with its owner. Exhibition content arrives as a spreadsheet that one person currently spends about three days hand-placing into a 100-page deck.
Content routes to one approved template and a library of 130 certified layout modules. If copy does not fit, the system chooses another layout or returns the content—it does not shrink the type until the problem disappears.
Every page clears both structural collision checks and a 200% rendered visual review. This module is in real-material validation.
PPT Studio · product user journey
User actionAI / tool actionHuman quality control / choice ⏸Automated quality control
Step 1 · Content side
Prepare materials
CommSphere · Pitch Deck Gen · Maven
Collect semi-finished decksDaily decks · Accelerate EN versions · outline copy — any of the three content tools
State the requirementBreakout first / Main Session limited · page count · output ratio
QC④ Human final review ⏸Owner + designer · humans stay accountable
Agency shifts to local fixesReduction measured by pilot, not assumed
Deliver to SCUsage tracking · feedback loop
Four quality gates + one two-way choice · no presumed savings numbers
Main path · PPT Studio · pilot Oct 2026
From the hub to an editable deck — seven clicks, two gates
User · click / decideAI · design workGate · governance
1Choose template
2To conversation
3Content gate
4Pick module
5Render → QA
6Two decisions
Editable PPTX
1
User
Open Design Center, click Choose template
Every path starts on this hub. The certified Dark 2026 card sits at the top; the chips underneath are the other ways in — Photo upload, Flash, Resize · 2027.
The template library opens with Dark 2026 ready to inspect.
Click
library opens · Dark 2026 ready
2
User
Preview the modules, click Continue to conversation
One certified visual system in the pilot: 130 editable modules, 19 layout types. The previews show what Dark 2026 contains before you commit.
A workspace conversation opens with the template locked in.
Click
workspace opens · template locked in
3
UserGate
Drop in your content file — the gate reads it first
draft-content.docx has no approval record, so it is held with a plain answer. Nothing renders from an unapproved draft.
Upload approved-content.docx and say what you need — the flow continues from the approved file only.
Chat · zoom
Held
Upload
approved file only — draft stays held
4
AIUser
Pick a module in the chat
Design MCP reads the 12 approved sections and recommends three fits for the opening. Click Open Pillars · T07 to confirm.
The editable deck starts building — 8 of 14 slides and counting.
Chat · zoom
Click
module confirmed · deck building 8/14
5
AIUser
Watch the render, then click Send to QA
Slides land in the Design Files panel as they finish — rendered, rendering, queued. The button stays greyed until every slide is in.
Send to QA unlocks only after render — click it and QA runs in the same thread.
Design files · zoom
Then click
14/14 rendered · QA runs in-thread
6
GateUser
Answer the two decisions on slide 7
QA checked all 14 slides; only slide 7 blocks. Click Apply fix for the logo safe zone. For the copy over budget, choose Switch layout — or return it to its owner.
The system never shrinks type to force a fit. Fit problems change the layout or go back to the content owner.
QA · zoom
Click
Or this
2 fixes cleared · ready for human review
Outcome · step 7
Open in PowerPoint → editable PPTX
Native PowerPoint objects, speaker notes kept, brand tokens intact. It opens in PowerPoint as a working file, not a flattened export.
Product 2
MUSE
Marketing User Interface & Style Engine — takes raw content and produces on-brand, production-ready email HTML, doing the job of both a designer and a front-end coder.
Why it exists
The channel team owns the full email lifecycle for 800K sellers — roughly 600 emails a year. Two steps had become the bottleneck: a designer or agency hand-coded every email in 6–8 hours, then a specialist worked through about 80 manual setup steps before it could be sent. Slow, an agency cost, and only a few people could do it.
MUSE takes the first job. A marketer uploads a plain Word doc; MUSE lays it out on-brand and writes cross-client production HTML in 30–60 seconds. No designer or vendor in the loop. Deployment is handed to ECHO, an RPA operator that turns the ~80 Pardot steps into two clicks.
The engine
Production-grade rather than a generic AI draft, because a senior design team's judgment is built into the engine and applied to every email.
MUSE + ECHO · EMAIL PRODUCTION JOURNEY
User actionTool actionHuman review / choiceAutomated check
Step 1
Prepare content
Approved Word documentThe marketer provides approved campaign copy.
Step 2
Generate with MUSE
Production HTML + QA report68 design modules · 3 templates · 94 coding rules · 40 brand guidelines.
Step 3
Approve creative
Human approvalReview the creative before deployment.
Step 4
Deploy with ECHO
Set up in PardotRPA operator · ~80 steps → 2 clicks.
Step 5
Send
Email delivered12.5 h manual → ~20 min.
Validation issue → return HTML or dynamic-field fixes to MUSE.
94 coding rules · on 68 modules / 3 templates
36Client compatibility — international & Chinese mail clients
16Responsive & performance
18Content generation rules
24Quality control & input validation
40 brand guidelines
13Colours & logo
9Typography
10Spacing & layout
8CTA hierarchy
Impact
Live on real seller campaigns since May, not a proof of concept. Announced to CNGS Marketing in August 2026.
6–8 h → ~1 minCreative + HTML, per email
4 days → 1.5 daysContent-to-send lead time
~7,300 hReleased per year, both products
¥200K → ¥0Annual agency fee, recurring
98 sends · 0 failures100% success since May
20 → ∞Email capacity — no longer limited by production
Design know-how now sits in the system rather than with one specialist, so any marketer can send a production email on their own.
MUSE in practice
A marketer starts with approved campaign copy in Word. MUSE turns it into production-ready HTML in 30–60 seconds, using a controlled module library instead of an open-ended visual generator.
Judgment stays with people. What leaves the queue is routine layout and coding, while creative approval and brand governance remain visible.
SYSTEM EVIDENCE · COMPLETE CAMPAIGN EMAILS · ORIGINAL CHINESE CONTENT
Captured from the supplied HTML files, including every section and footer. Template variables are shown as supplied; the newsletter’s recipient-specific QR image requires send-time data.
MUSE 2 · Hyper-personalization
MUSE 1 produced one on-brand email per campaign. MUSE 2 produces one email per seller, with content written by an AI engine every business day. Live since September 2026, it connects MUSE to MAVEN, the seller-launch personalization program.
Problem · Agent-to-agent handoff
The biggest change was upstream: the content creator moved from a person to an agent. In MUSE 1, people prepared content in Word, where headings, paragraphs, lists, and emphasis already carried the intended structure. In MUSE 2, the content agent writes each seller’s email into one row of a sending list. The challenge was therefore agent-to-agent handoff: how could the design agent pass its email modules to the content agent through a plain-text spreadsheet cell, reliably across thousands of rows?
Solution · A compact semantic contract
I solved this with a compact semantic contract: seven tags for title, preview, salutation, body, sub-heading, list, and button, plus two inline markers for emphasis. I packaged the contract as a skill for the upstream content agent, enabling it to generate seller-specific content in a structure that maps directly to MUSE’s design modules. A deterministic parser then renders every row using the same brand and layout rules. No model runs at send time, so the output remains consistent while the content stays personal to each seller.
LIVE · HYPER INTERFACE2,189HTML outputs · one workbook batch
BUILT-IN QUALITY CONTROL
Green QC. Checked and ready.
The green status opens the checks behind the selected email’s QC result.
010203
01 · Batch navigator
Viewing output 1 of 2,189
03 · Green QC statusPassed for the selected email
02 · QC panel, enlarged
69 text groups19 number groups4 percentage groups2 URL groups
Also checks dates, email addresses, amounts and dynamic placeholders.
LIVE HYPER INTERFACE · 2,189 HTML outputs in one batch. Green QC shows the selected email has passed its checks.
SOURCE WORKBOOK
One content cell → one email.
Workbook overview · column G highlighted
contentColumn G · detail
The source file carries each email’s tagged structure in the content column. Original Chinese copy is preserved.CONTENT CONTRACT · TYPESET FROM THE AUGUST 2026 SPECIFICATION
Quality control
We built an independent QC step into the live pipeline. It checks that every rendered email faithfully preserves its source content before it can be sent.
HYPER · QUALITY CONTROL JOURNEY
User actionTool actionHuman review / choiceAutomated check
Step 1
Source content
Provide the sourceOne tagged content cell per email.
Step 2
Generate HTML
Assemble the emailUse the shared template and source content.
Step 3
Check fidelity
Compare against sourceComplete · accurate · correctly structured.
Step 4
Resolve the result
Pass → ready to sendGreen QC status appears in the interface.
Fail → retry and recheckUp to 3 retries.
Still failing → human reviewReview the reason and decide the next action.
QC checks fidelity. The content owner remains responsible for source accuracy and compliance.
Impact
MUSE 2 is the piece that turned MAVEN's email channel from a concept into a daily send. MAVEN can write per-seller content; without a deterministic HTML generator, no row of it becomes a brand email a seller can receive.
+55% capacity+22 sends a month on the 40 quota, zero added designer hours
Flat QC costReview by exception. Volume grows, review effort does not
No drift at scaleConsistent rendering rules across every row. Zero LLM in the send path
Program result · MAVEN pilot, first month
3.2×click rate vs lifecycle baseline (11.4% vs 3.6%) across 2,558 prospects reached
Program-wide figure for the personalization program across WeChat, WeCom and email, reported to WWGS SLT. Not attributable to the rendering pipeline alone.
Quality holds by rule rather than by a designer watching. Next: the same contract extends to tables, images and per-seller links.
What's next
Phase 1 · live
Standard ad-hoc + dynamic-content email. One upload generates multiple seller-segment versions automatically.
Phase 2 · live Sep 2026
Hyper-personalization. One individually generated email per seller, one batch every business day.
Next
Extend the content contract to tables, images and per-seller links.
Later
Decentralise. Move from one shared machine to a server so every marketing POC can run the AI directly.
View complete email ↗
View complete email ↗
View complete email ↗
