01 / My Evolving Approach to AI Design
From input-output frameworks in Shimo AI to task guidance in Zoom AI Sheets and access boundaries and verification in my own tools, I explore how open-ended capabilities can fit into clear, inspectable workflows.
Explore the storyFOUR QUESTIONS THROUGHOUT THE WORK
What is the person trying to accomplish, and what context does the system have?
When should the capability appear, and how does the result return to the work?
What can the system access or change, and how can the person intervene?
What supports the result, and how can errors be found and handled?
01 / Shimo AI · 2023
In Shimo AI, I began with a question: how should a product organize capabilities when people’s content needs and ways of expressing them keep changing? Preset prompts can help people get started, but are not enough to structure the whole experience. I organized the framework around user inputs and AI outputs, then mapped interactions across text, images, files, audio, and video.
Task shortcuts within an extensible framework
For image generation, the exploration broke common descriptions into selectable aspects and keywords, helping people articulate their intent without first learning to write a complete prompt.
The AI panel was envisioned as a transfer point between products.
The result becomes something people can keep working on.
02 / Zoom AI Sheets · 2025–2026
For Zoom AI Sheets onboarding, I organized demonstrations around data preparation, analysis, formulas, and cross-sheet work, using role and behavior to select relevant examples. The proposal lets people skip guidance and explore when ready, using a separate demo sheet to protect their real work.
Use role and behavior as context.
Allow people to skip and return.
Keep the demo separate from actual data.
03 / Context, capability & collaboration
A useful AI experience extends beyond the answer box. In these Zoom AI Sheets feature animations, I communicate how people bring in meeting context, clean existing data, request analysis, inspect a formula and return to collaboration. These are five related scenarios, not a claim that one autonomous agent completes them all.
A sheet already contains actual values and targets. The task is to highlight actual values below their row target. I use this scenario to connect intent, data boundaries, preview, confirmation and recovery across the following chapters.
“Mark actual values below their row target in red.” Keep the selected cells as context.
For the current-selection example, read B2:C3 and format B2:B3. Ask for a source when it is unclear.
Preview =B2<C2, the destination range and red formatting together. Keep business values unchanged.
Let people apply, adjust or cancel after reviewing the suggestion.
Keep the result as a native rule that can be edited or undone. Offer manual creation when generation fails.
ZOOM AI SHEETS / 01
Select the meeting as a source, describe the task, then see a structured sheet.
The animation connects the meeting source, the request and the resulting sheet, making the origin of the work visible.
Meeting context → a sheet people can keep editing
Original feature demonstration used to communicate the interaction flow.
ZOOM AI SHEETS / 02
Open the skills picker, choose Clean Data and inspect the changes in the sheet.
I kept the entry, selected skill and result in one short sequence, so the capability reads as a task rather than a prompt-writing exercise.
Choose a skill → process the current data → inspect
Original feature demonstration used to communicate the interaction flow.
ZOOM AI SHEETS / 03
Select cells and invoke Analyze for insights from the existing context menu.
The animation shows the handoff from a familiar spreadsheet action to the AI panel. The selected range remains the key context that the design must preserve.
Selection → contextual entry → analysis panel
The analysis reply uses placeholder sales figures that do not match the equipment sheet shown earlier. This clip illustrates the entry and response sequence; its percentages are not a verified analysis or project result.
ZOOM AI SHEETS / 04
Describe the calculation beside the destination cell, inspect the generated formula and accept or reject it.
I highlighted the referenced cells, the waiting state with Cancel, and the formula explanation with accept/reject controls. Users see both what is proposed and where it will go.
Intent → wait / cancel → formula → accept / reject
Original feature demonstration used to communicate the interaction flow.
ZOOM AI SHEETS / 05
Choose a file in the meeting share flow, then enter the sheet with participant and follow-mode cues.
The project records identify my role in aligning Meeting Sharing across Sheets and Paper. The animation makes that handoff visible through file selection, sharing and the collaborative workspace.
Select a file → share → review together
Original feature demonstration used to communicate the interaction flow.
CHOOSE THE LEVEL OF AI INVOLVEMENT
The AI-native formula proposal makes references and outputs part of a familiar cell calculation.
A chart or analysis request can return a bounded suggestion that people preview or reject.
Extraction, matching and reshaping need a visible plan and stepwise confirmation in the onboarding proposal.
The three modes come from the onboarding specification. They describe how the proposed capabilities are organized, not a single released autonomous workflow.
04 / Spreadsheet interaction design
As a spreadsheet interaction designer, I focus on keeping people and AI aligned throughout the conversation: which sheet and cells provide the data, and where the result will take effect. My design principle is to keep these references and action scopes visible while people express intent, review suggestions, and confirm changes, so formulas, charts, and formatting remain grounded in the data people provide and can be checked against their source cells.
“Mark actual values below their row target in red.”
B2:C3Apply to B2:B3| A | B | C | |
|---|---|---|---|
| 1 | Month | Actual | Target |
| 2 | Jan | 82 | 100 |
| 3 | Feb | 112 | 100 |
| 4 | Mar | 91 | 100 |
| 5 | Apr | 106 | 100 |
Read actual values in column B and the existing targets in column C, row by row.
Apply formatting only to these cells. Keep their business values unchanged.
=B2<C2Check references and scope → preview → confirm
If the source is missing or the reference is unclear, ask people to select or provide the data first. AI suggestions should carry a traceable source and an explicit destination.
05 / Within familiar spreadsheet actions
People already have familiar spreadsheet workflows: selecting data, creating charts, entering formulas, and setting conditional formatting. I want AI to appear within these actions, using the current object and selection as context. People can describe a change in a chart, explain a calculation while writing a formula, or state what they want to highlight while setting a rule, then inspect and refine the result through familiar controls.
AI should support the task already in progress. Its entry point, context, and resulting changes should belong to the same workflow.
A lightweight input within chart editing uses the current data and settings. A preview leads back to the same chart and its existing controls.
“Show monthly sales trends, grouped by region.”
Original action → express intent → preview → keep editing
An entry near cell or formula editing turns a calculation goal into a reviewable formula, with explicit references and a destination cell.
“Calculate each row’s share of total sales.”
Original action → express intent → preview → keep editing
An entry within rule creation turns the desired effect into an inspectable condition, range, and style. The result remains a standard editable rule.
“Mark sales below their row target in red.”
Original action → express intent → preview → keep editing
The original design fragments locate the entry points. The examples and preview-to-edit workflow are narrative design proposals; release status is not asserted. The sidebar remains available for open-ended and cross-object work.
06 / Explanation, native editing & fallback
Generation is only one step. I reorganized chart settings so people can distinguish the data being used from the way it is displayed. In the formatting-rule proposal, I carried the same principle further: show the condition, range and style before applying, then keep the result in the native editor with undo and manual fallback.
The PM considered placing a color-explanation tooltip beside the cell, so people could understand why it had a particular color.
The adjacent explanation would obstruct ordinary browsing. An explanation needed in a specific situation would occupy the everyday reading space.
We dropped the cell-adjacent tooltip option. I argued for keeping routine browsing clear when providing explanations for special states.
My principle: everyday use should not give up its space or usability to accommodate a special state. Explanation belongs at the moment it is needed.
Source range · series · names
Title · labels · axes · grid
REVIEW BEFORE APPLYING
=B2<C2B2:B3Concept illustration based on chart review records and the formatting-rule proposal. It does not reproduce a released screen.
Separate data configuration from presentation. Show the references, condition and destination rather than only a finished chart or color change.
Generated formulas and rules should remain standard editable objects. AI supports creation; familiar controls support precise adjustment.
The proposal retains undo after applying and manual creation if AI fails or is unavailable. Keep the task context instead of making people start over.
DESIGN THE EXCEPTIONS, TOO
Keep the request and selected range → return to manual rule creation.
Retain the confirmed destination; never silently expand or replace the scope.
Preserve ordinary spreadsheet tools and respect the existing permission and usage rules.
Interaction requirements from the rule proposal, pending product and engineering review.
07 / Recent independent practice
In independent tools, I continued examining the decisions behind the interface. A2A Contract Hub explores access and delivery: who can receive which information, and how decisions are recorded. The RSU cost basis reconciler focuses on deterministic calculations and reviewable workpapers.
Purpose & requested information
Allow or deny within explicit rules
Decision basis & delivery receipt
One meeting record, two task-ready views: the internal view retains an owner email; the external view retains the action and due date, with the owner unassigned.
View source on GitHubLocal alpha · synthetic meeting demo · no live task-platform integration
| Proceeds | $12,000 |
|---|---|
| Reported basis | $0 |
| Compensation basis | $10,000 |
| Reconciled gain | $2,000 |
12,000 − 10,000 = 2,000
I turn input records into deterministic calculations and reviewable workpapers. The complete built-in path currently uses a synthetic case.
View source on GitHubLocal-first prototype · real broker-PDF flow remains unverified
08 / Cross-product delivery & validation
An AI interaction still has to work within a real product: file and sheet permissions, long translated labels, shared edits and failed requests. I use interaction specifications, key screens, feature animations and review with product and engineering to clarify those boundaries. I also define what to observe next, without treating a design proposal as measured success.
I aligned Meeting Sharing across Sheets and Paper. File selection, sharing and the collaborative workspace should feel like one handoff, even when several products are involved.
Chart reviews include long translated labels and adjustable panel widths. Permission work distinguishes file-level toolbar states from sheet-level rules to avoid a shifting interface when people switch sheets.
I connect entry conditions, range behavior and exception paths to key screens and motion examples, then review them with product and engineering. A clear acceptance case matters as much as the visual.
Observe task selection, skipping, importing and voluntarily opening the demo. Separate discovery from completion.
Ask people to locate a reference and the cells that will change; follow the path from the rule explanation to editing.
Cover apply, undo, failed generation, changed selections and shared-rule conflicts with explicit acceptance cases.
Intent · workflow · control · verification
I aim to design a complete process people can understand, use and inspect.
MY CONTRIBUTIONS & EVIDENCE
| Practice | My contribution | Evidence on this page |
|---|---|---|
| Shimo AI · 2023 | Interaction framework & intent articulation | Input/output analysis, keyword selection & editable-output concepts |
| Zoom AI Sheets · onboarding | Task-based onboarding & motion production | Task selection designs, onboarding specification & five original animations |
| Zoom AI Sheets · core interaction | Chart settings, rule ranges & meeting-sharing alignment | Original design fragments, review records & rule proposals |
| A2A Contract Hub | Independent product design & implementation | Data-delivery boundaries & task-ready views · local alphaView source on GitHub |
| RSU Basis Reconciler | Independent product design & implementation | Deterministic calculation & reviewable workpapers · synthetic previewView source on GitHub |
Team-product contributions and independent implementations are distinguished. Original animations and design proposals do not establish a release status; repository access follows existing permissions.