Concept | AI-native prosumer productivity (agentic, human-in-the-loop)

Understudy

Understudy

Understudy

An AI chief-of-staff that drafts and stages your work across your tools, then waits for one-tap approval before it acts.

An AI chief-of-staff that drafts and stages your work across your tools, then waits for one-tap approval before it acts.

An AI chief-of-staff that drafts and stages your work across your tools, then waits for one-tap approval before it acts.

Role

Concept | Product design | AI interaction

Timeline

2026 | Self-directed concept

Stack

Space Grotesk + Inter + IBM Plex Mono | single amber accent

Platform

Web + mobile

Context

A self-directed concept built to close a specific hole in my portfolio: every existing piece is web, desktop, or print and analytical or brand-led. Nothing shows the frontier skill hiring managers screen hardest for in 2026, designing trust, provenance, and human control for a generative, agentic system. Understudy is that piece, framed honestly as a concept with no client and no invented outcome metrics.

Problem

Agentic assistants split into two failure modes. One over-automates: it acts on your behalf, quietly, and you find out after the fact, so trust collapses on the first mistake. The other under-designs: it dumps raw model output into a chat window with no sources, no confidence, and no way to steer, so the human stays the bottleneck. The real design problem is to give a solo operator genuine leverage while keeping a visible, revocable gate between the model’s intent and the real world, and to make the model’s work legible enough to approve in seconds, not minutes.

Approach

I designed around a single load-bearing idea: staging. Understudy never touches the outside world by default. It drafts, it plans, it stages, and then it stops at an explicit approve or reject gate. Four systems carry that idea. A plan-then-act view exposes the agent’s steps as legible, interruptible text instead of a spinner. Provenance is treated as receipts, every claim the agent makes carries a source chip you can open. Confidence is a calibrated three-tier signal, never a false-precision percentage. And a single electric amber accent is rationed to mean exactly one thing, this needs your hand, so the eye travels only to where a decision is actually owed. Kinetic typography carries the state change from Drafting to Staged to Sent, so motion reads as the system thinking rather than as decoration.

Outcome

A concept case study, not a shipped product. The deliverable is a coherent ‘calm command’ design world across responsive web and a companion mobile surface, expressed as four key screens with full build specs. Its job in the portfolio is evidence of judgment on the hard AI-UX problems, plan transparency, provenance, calibrated confidence, editable drafts, undo, and an explicit human gate, in a light, architectural register that is visibly distinct from both the dark chrome of the rest of the roster and the warm, earthy register of the companion concept.

Decisions | Reasoning

Make staging the product's physical law: nothing acts until approved, and the default toggle ships set to Staged, not Autopilot.

Make staging the product's physical law: nothing acts until approved, and the default toggle ships set to Staged, not Autopilot.

Responsible, human-in-the-loop design has to be structural, not a setting buried three levels deep. Putting the gate in the center of the model and making the safe mode the default is the clearest possible signal that the human holds the trigger. It also turns the scariest property of agents, that they act, into a designed, observable moment the user owns.

Responsible, human-in-the-loop design has to be structural, not a setting buried three levels deep. Putting the gate in the center of the model and making the safe mode the default is the clearest possible signal that the human holds the trigger. It also turns the scariest property of agents, that they act, into a designed, observable moment the user owns.

Ration the single amber accent to one meaning only: 'a decision is owed here.' Everything the agent has not touched or has already completed stays neutral.

Ration the single amber accent to one meaning only: 'a decision is owed here.' Everything the agent has not touched or has already completed stays neutral.

If color decorates, it stops signaling. By spending amber exclusively on the approval surface (staged cards, the gate bar, the pulsing current step), the interface becomes navigable by color alone: a user scanning on their phone knows in under a second where their attention is required. Completed work resolves to a calm positive, and untouched work is quiet, so the queue never feels like an alarm.

If color decorates, it stops signaling. By spending amber exclusively on the approval surface (staged cards, the gate bar, the pulsing current step), the interface becomes navigable by color alone: a user scanning on their phone knows in under a second where their attention is required. Completed work resolves to a calm positive, and untouched work is quiet, so the queue never feels like an alarm.

Treat provenance as receipts, not footnotes. Every drafted fact carries an openable source chip in monospace, and each staged action lists its sources inline.

Treat provenance as receipts, not footnotes. Every drafted fact carries an openable source chip in monospace, and each staged action lists its sources inline.

You cannot responsibly approve what you cannot verify, and verification has to be faster than redoing the work yourself. Rendering sources as compact, technical, tabular chips makes them read as system-generated evidence rather than prose, which both builds trust and lets the user spot-check a claim without leaving the card. The monospace family is doing semantic work: it marks 'this is machine truth you can audit.'

You cannot responsibly approve what you cannot verify, and verification has to be faster than redoing the work yourself. Rendering sources as compact, technical, tabular chips makes them read as system-generated evidence rather than prose, which both builds trust and lets the user spot-check a claim without leaving the card. The monospace family is doing semantic work: it marks 'this is machine truth you can audit.'

Show confidence as a calibrated three-tier verbal-plus-bar signal (High / Worth a look / Uncertain), never a percentage.

Show confidence as a calibrated three-tier verbal-plus-bar signal (High / Worth a look / Uncertain), never a percentage.

A number like '87%' invents precision the model does not have and quietly trains the user to over-trust it. A three-tier signal is honest about resolution, maps cleanly to a decision (approve fast, glance, or verify), and lets me pair the lowest tier with an explicit caution note that names why the agent is unsure. This is 'responsible design' made legible at the exact moment of the decision.

A number like '87%' invents precision the model does not have and quietly trains the user to over-trust it. A three-tier signal is honest about resolution, maps cleanly to a decision (approve fast, glance, or verify), and lets me pair the lowest tier with an explicit caution note that names why the agent is unsure. This is 'responsible design' made legible at the exact moment of the decision.

Render the agent's thinking as legible, interruptible plan text with a streaming caret, not a loading spinner, and let the user stop it mid-plan.

Render the agent's thinking as legible, interruptible plan text with a streaming caret, not a loading spinner, and let the user stop it mid-plan.

A spinner hides the one thing that builds or breaks trust in an agent: what it is about to do and why. Streaming the plan as readable steps makes the reasoning auditable in real time and, crucially, interruptible, so the human can correct course before an action is even staged. Interruptibility is the difference between a tool that assists and one that runs away from you.

A spinner hides the one thing that builds or breaks trust in an agent: what it is about to do and why. Streaming the plan as readable steps makes the reasoning auditable in real time and, crucially, interruptible, so the human can correct course before an action is even staged. Interruptibility is the difference between a tool that assists and one that runs away from you.

Use kinetic typography to carry state transitions (a verb morph from Drafting to Staged to Sent) rather than icon swaps or toasts.

Use kinetic typography to carry state transitions (a verb morph from Drafting to Staged to Sent) rather than icon swaps or toasts.

State is the most important information in an agentic UI and it changes constantly. Animating the state word itself puts the change where the user is already reading and makes motion communicative instead of ornamental. It also gives the product a distinct expressive voice that is precise rather than playful, reinforcing the 'calm command' register and setting it apart from the two other design worlds in the portfolio.

State is the most important information in an agentic UI and it changes constantly. Animating the state word itself puts the change where the user is already reading and makes motion communicative instead of ornamental. It also gives the product a distinct expressive voice that is precise rather than playful, reinforcing the 'calm command' register and setting it apart from the two other design worlds in the portfolio.

Results

A concept, not a shipped product: four full-spec screens built to demonstrate judgment on hard AI-UX problems like transparency, confidence, and human control.

Next project

Wend

Wend