Nudge

01 · Idea

Open the prototype (opens in a new tab)

Case study · 2026

Never miss what your coding agent needs.

A phone app for people who run Claude Code, Codex and Cursor. When an agent stops to ask, Nudge explains it in plain words and lets you answer safely, from anywhere.

  • Product Designer · solo
  • Concept + live prototype
  • iOS & Android
  • 2026
Open the prototype (opens in a new tab) Read the story
A Claude Code chat for my-app: the agent explains its work, then asks to run rm -rf dist to delete the old build folder, marked low risk, with Deny and Approve.
Nudge Home: Windows PC awake; 4 waiting, 1 working, 2 completed; under Needs you, Approve all for 3 low-risk requests and a Claude Code card asking to delete the old build folder.

02Key screens

  • Home: Needs you comes first, with Approve all for 3 low-risk requests and a Claude Code card to delete the old build folder.
    HomeWhat needs you comes first.
  • Approval, 1 of 3: Low risk, Delete the old build folder, the command rm -rf dist, what this does in plain words, what the agent did before, and Deny and Approve.
    ApprovalPlain words and a risk level.
  • A high-risk request in the chat: Force-push to main, git push --force origin main, with Deny and Hold to approve, and a warning that other people’s recent changes could be lost.
    Hold to approveHigh risk is never one tap.
  • A Claude Code chat: your message on the right, the agent’s replies as plain text, tool activity as one line, and an approval card inside the chat.
    ChatWorks like Claude and ChatGPT.
  • Tasks: every task from Claude Code, Codex and Cursor in one list, filtered by agent and grouped into Needs you, Running and Done today.
    TasksEvery agent in one list.

03Problem statement

  1. 01

    Work stops

    The agent waits until the person comes back.

  2. 02

    The session can drop

    On Windows laptops, sleep drops the connection.

  3. 03

    Rushed decisions

    People approve without reading; work is redone.

  4. 04

    Juggling agents

    No single place shows which one is waiting.

The agent doesn’t fail. It waits.

And nobody’s there to answer. People who run AI coding agents, from professional developers to non-coders, step away while the agent works: a short break, a class or meeting, an overnight run. While they’re away, the agent stops to ask for permission or a question, and they don’t see it.

How might we make sure people never miss what their coding agents need, and can answer correctly and safely from wherever they are?

04Research

How it connectsEnd-to-end encrypted
  1. Coding agentasks permissionon your computer
  2. Helpercatches requestson your computer
  3. Relaypasses sealed messageson the internet
  4. Nudgeyou decidein your pocket

Why a relay. A phone can’t reach a laptop directly on most networks, and the relay needs no setup.

One approval, step by step
  1. Agent stops to ask permission
  2. Hook hands the request to the helper
  3. Helper sends it to the relay
  4. Relay pushes it to your phone
  5. You tap Approve
  6. The answer returns via the relay
  7. Helper tells the agent: allow

Two checks before any design

Can it be built, and is the problem real? A great design for an impossible or imaginary product is wasted work.

Can it be built?

Agents

Only what each agent supports

A button that can’t work is worse than no button.

ApprovalsQuestionsControl
Claude Code
Codex
Cursor
Encryption

Encrypted, paired by QR

A phone that can approve commands is a remote control for a computer.

The relay carries locked messages it can’t read.

Cost

Free tier

Cost checked first

No big server bill. The real fixed costs are the app store accounts.

Undo

Undo restores files only

Saying so up front avoids false confidence.

  • Files
  • Commands already run
Technical

Constraints mapped first

Each one, with its solution, before drawing a screen: a design decision in disguise.

  • Phone offline“Not sent” · Retry
  • Computer sleepsKeep-awake · banner
  • Request times outBack to the computer

Is the problem real?

The brief I gave Claude

Find, using facts only, whether this problem is real: what developers report when they leave agents running and step away.

  • Facts onlyso I wouldn’t build on a hunch
  • Separate columnsso my hypotheses couldn’t pass as evidence
  • Only features the facts supportso every feature traces back to a real problem
  1. Fact

    Agents stall for long stretches on one permission

    I assumed: single-agent users are covered by official apps

    The core problem is real

  2. Fact

    Official phone apps already handle basic approvals

     

    Nudge must solve what’s missing

  3. Fact

    Laptops sleeping drop the session

    I assumed: maybe the biggest unsolved pain

    Keep-awake + a clear “offline” state

  4. Fact

    Approving on the move = approving quickly

    I assumed: nobody can judge a command in seconds

    Plain words + risk level

  5. Fact

    A phone approval once never reached the computer, with no error

    I assumed: one silent failure breaks trust

    Sent → Reached → Running

  6. Fact

    Each official app shows only its own agent

    I assumed: the pain sits with multi-agent users

    One place for every agent

Facts before features

I didn’t want to design for a problem I’d only imagined, so I tested the idea against facts first.

What people told me

14

People who use coding agents, heard before building

The facts said the problem existed, but online posts show what’s loud, not always what’s common. They face this problem, and said an app like this would help them.

05Concept thinking

Feature map · a hub, a TV remote

Setup

3
  • Pair by QR
  • See and remove paired phones
  • Claude Code, Codex, Cursor

Home (the hub)

3
  • All sessions on one screen
  • One-line live status
  • Computer status + keep-awake

Approvals and questions

6
  • Approve / deny from the lock screen
  • Plain explanation + risk
  • Summary of what the agent did
  • Answer questions
  • Delivery confirmation
  • Timeout back to the computer

Control

4
  • Stop → Continue / Redirect / Undo
  • Model and intensity
  • Voice
  • Start a new task questioned, then kept

Later

2
  • Approval history
  • Trust rules

Need real usage first.

Cut

2
  • Commit and push buttonsApprovals already cover them
  • Editing code on the phoneClutters a quick tool

A remote for all your agents

Not just a way to say yes. One hub that shows what’s working, waiting and done, because people running agents need the whole picture, not one prompt at a time.

  • PrincipleBuilt on the gapsBasic alerts and approvals already exist. The gaps were one place for every agent, plain-language risk, proof an answer arrived and a computer that stays awake.
  • PrincipleEvery feature answers a factFrom pairing by QR to voice replies, each kept feature traces back to the research.
  • KeptNew task: questioned, then keptA task often finishes while you’re away, and you want to start the next step without going back.
  • CutEditing code and commit buttonsA phone is the wrong place to write code, and commits already arrive as approvals.
  • LaterHistory and trust rules“Always allow this in this project” needs real usage before it can be designed well.
Concept sketches
  1. 1 · Setup
    Connect your computerScan code shown on computerAgents foundClaude CodeCodexCursorSettings › paired phones
  2. 2 · Home (the hub)
    Your computerAwakeClaude Code · appWaiting for youCodex · apiEditing login.js…WorkingCursor · siteDone+ New taskVoice
  3. 3 · Lock screen notification
    Claude Code · appWants to delete old build filesLow riskApproveDenyOpen
  4. 4 · Approval detail
    What it did so farIt wants to runrm -rf distplain explanationLow riskApproveDenySent → reached computer → running
  5. 5 · Answer a question
    Which login method should I add?Email and passwordGoogle sign-inType your answerVoice
  6. 6 · Session control
    Codex · apiEditing login.js…Model: defaultIntensity: mediumStopAfter stopContinueRedirectUndo

06User flow

Approvals · main path left, edge cases right
  1. Agent asks permission
    • Several requestsGrouped per session, queue
  2. Notify: explanation + risk
    • High riskFull command shown
    • Answered on computerNotification removed
    • No answer in 10 minGoes back to computer
  3. (optional) Open full detail
    • AI explanation offRisk and command only
  4. Approve (fingerprint) or Deny
    • Fingerprint failsUse phone PIN
    • Tapped too lateExplain + message agent
  5. Answer sent
    • Phone offlineNot sent, retry
    • Didn’t reach computerShow failure, retry
  6. Reached computer, running

Every path, including failures

I mapped every path from sign-in to every feature, including what happens when things go wrong, because that’s where trust is won or lost.

Set up

  • Familiar sign-inGoogle or email and password, the two most familiar options.
  • A short explanation firstThe first step happens on the computer, so people need to know why before they install anything.
  • You confirm each agentThe helper finds agents automatically, and nothing connects without a yes.
  • Notifications, with a reasonAsked for right after pairing, the moment the value is obvious.
  • Keep-awake from the startOffered as soon as Home opens, so the session doesn’t drop on the first use.

Answer

  • Face ID or fingerprintApprove from anywhere. Typing needs the app open, because a typed answer deserves a full screen.
  • One queue per sessionSeveral requests are reviewed in order instead of hunting.

Control

  • Stop confirms only when files changedAsking every time teaches people to tap through confirmations.
  • Undo, one step at a time“Go back further” is there too, so bigger rollbacks are a deliberate choice.
  • Limits explained, not hiddenCursor’s Stop button is shown with an explanation, so people understand the limit instead of wondering where it went.
  • Take over from StopTapping Stop asks once, “This stops it in your terminal and continues from your phone”, instead of a Take over button.

When it fails

  • Nothing fails silentlyEvery failure gets one clear message and one next step, from a sleeping computer to a failed QR scan.

07Flowcharts

Onboarding
While work runsBackground alerts
Home (the hub)
SettingsPhones, computers, alerts
  • Approvals
  • Questions
  • Control
  • New task
  1. Install the app
  2. Sign in (Google or email)
    • Forgot passwordReset link
  3. Read short setup steps
  4. Tap “Email me the link”
    • Email not receivedResend
  5. Install helper on computer
    • Install failsRetry / other method
  6. Scan QR code with phone
    • Scan failsRetry or type 6-digit code
    • Code expiredHelper shows a new code
  7. Confirm detected agents
    • No agents foundInstall links, scan again
  8. Connect each agent (Yes)
    • User taps NoShown as not connected
  9. Choose keep-awake
    • Keep-awake declinedChange later
  10. Allow notifications
    • Notifications deniedWarning on Home, Turn on
  11. Home
  1. Open app
    • Notifications offWarning + Turn on
  2. Computer switcher + status
    • Computer offlineNotify once + banner
    • Agent disconnectedOne-tap reconnect
  3. Waiting for you, first
    • No sessions yetEmpty state + New task
  4. Working · one-line status
  5. Done · clears in 24 h
  6. Tap session or New task
  1. Agent asks a question
  2. Notification shows it
    • Up to 3 short optionsTap one + fingerprint
  3. Open answer screen
    • Plan to reviewSummary + read full plan
  4. Pick, type or speak
    • Empty answer“Type an answer first”
  5. Voice: check and edit text
    • Words misheardFix before sending
  6. Sent, reached computer
    • Timeout / offline / too lateSame as approvals
  1. Open a session
  2. Status, model, intensity
  3. Tap Stop
    • Agent can’t do thisButton shown, explains why
  4. Files changed: confirm
    • No files changedStops right away
  5. Session stopped
  6. Continue, Redirect or Undo
    • UndoLast step, or go back further

before feedback New task, as first designed

  1. Tap New task
  2. Choose computer
    • Computer offlineCan’t start, explains why
  3. Choose agent
    • Agent can’t start tasksShown, explains why
  4. Folder: recent + search
    • Search finds nothingNo project found
  5. Type or speak instruction
  6. Safety: ask every action
    • Change safetyBefore Start
  7. Start, appears in Working
  • Usage limit reached
    Paused, continues at 3:45Session line only, no alert
  • Task finished
    NotificationCan turn off
  • Task failed
    NotificationCan turn off
  • Session done
    Fades, gone after 24 h

08Wireframes

  • Wireframe of an approval, 1 of 3: what it did so far, it wants to run rm -rf dist, what this does, Low risk, and Deny and Approve.
    Approval
  • Wireframe of the computer helper: Pair your phone, with a QR code and a 6-digit code to type if scanning fails.
    Computer helper

53 wireframes before any style

Every screen in the flow at mid-fidelity, so structure and flow were settled before style could distract from them.

Mid-fidelity wireframe of Home: Windows PC · Awake at the top, Waiting for you with Claude Code · my-app and Codex · api, then Working and Done, above a floating tab bar with four items.
  1. Floating bottom navigationLike iOS apps use today. Content stays visible, and it’s reachable with one thumb.
  2. Four items, not threeThree looked unfinished at the bottom of the screen.
  3. Add a computer, in the dropdownThat’s where people look for their computers.
  4. One more checkFor anything missing before moving on, because a missing screen costs minutes now and days after visual design.
  5. Feasibility, checked againAgainst an app built on the same idea, so the design wouldn’t promise anything impossible. It confirmed Undo restores files only, Cursor can’t be controlled, and how to handle sessions started in the terminal.

09Moodboard

  • Liquid glass direction: a lock-screen notification with glassy Approve and Deny buttons over a vivid purple and orange gradient.
    Liquid glass Rejected
  • Neon accent direction: a black Home with a bright lime Waiting tile.
    Neon accent Rejected
  • Warm editorial direction: Home on a warm cream and peach gradient.
    Warm editorial Rejected
  • The chosen minimal base: a light Home with a black Waiting tile, white cards and a floating tab bar.
    Chosen base: minimal

Most directions, rejected on purpose

I explored colour and type directions with Claude, because the first options were rarely good enough to build a product on.

  1. RejectedDated styles rejectedThe first set looked like styles from 2020, and a tool for AI agents has to feel current.
  2. Asked forA mix of two directionsHeld to HCI and Google guidelines, contrast and spacing, so a style choice couldn’t turn into a usability problem later.
  3. RejectedCorrect isn’t enoughDirections that looked lacking were rejected even when they followed the rules. It has to feel premium.
  4. ChosenCalm minimalismModern effects and strong typography, because a screen you decide on in seconds has to stay calm. Dribbble references set the quality bar.

10Design system

Type · IBM Plex

  • Large title 28
  • Title 22
  • Headline 17
  • Body 17
  • Footnote 13
  • rm -rf dist · Mono 15

Mono only for commands, project names and times, so code reads as code.

Colour · black & white, colour = risk

  • Background#0E0F12
  • Surfacewhite 6.5%
  • Ink#F5F5F7
  • Secondary#A1A1A6
  • Low risk
  • Medium risk
  • High risk
  • Working

Colour means risk or status, and never alone: always an icon and a word.

Components · 44 pt targets

DenyApprove
1Email and password
2Google sign-in
Reply to Claude Code…
AI explanation

Numbered options first, typed answer last, so every question works the same way.

One decision at a time

Each option checked against platform guidelines, so no visual choice could break usability later.

  • Colour came lastSo it couldn’t hide weak layout or hierarchy.
  • 8-point grid, 44 pt targets16px side margins keep spacing consistent, and every tap lands.
  • Glass only on bars and buttonsNever as decoration, so content behind turns to haze instead of fighting the labels.
  • Apple’s type rampNothing below 12px, so text stays legible and the app feels native.

11Design

  1. The first full design: a light Home titled [App name] with a black Waiting tile, a request card with Deny and Approve, and a question card with numbered options.
    Before: first design
  2. The design on a grey background, which was rejected.
    Grey background Rejected
  3. The final dark Home: Needs you first, Approve all for 3 low-risk requests, and a Claude Code card to delete the old build folder.
    After: final
  • In-app mark
  • App icon (dark tile)
  • App icon (light tile)
  • NudgeLockup

Better by removing things

The first full design got a hard review. Most changes removed things, because every extra element competes with the one decision on screen.

  • Needs you, firstHome keeps Needs you → Working → Completed, because what needs you must always come first.
  • New task in the centreSettings moved to the top-right and New task to the centre of the tab bar, where the most-used action belongs.
  • No countdownsA ticking clock pushes people to approve without reading. Requests show “x min ago” and “Goes back to your computer after 10 min” instead.
  • Friction matches riskLow and medium: Approve, then Face ID or fingerprint. High: Review, then Hold to approve, because a dangerous command deserves a deliberate action.
  • Colour on the icon onlyAlerts never colour the text, so warnings stay readable and the screen stays calm.
  • Black and white summariesSummary boxes use icons, because colour is reserved for risk.
  • No grey backgroundIt made the whole app look unfinished.
  • Named NudgeA logo of rings opening outward, a signal reaching you. Blue was dropped for black and white, which leaves colour free to mean risk.
  • Dark first, light tooDark is the version I refined furthest in Figma. Every component exists in both modes, so either stays consistent.

12Prototype

The live prototype’s Home screen.
  • A new Claude Code chat: What should Claude Code do?, the last used project as one line you can tap to change, three suggestions, and the message box.
  • A high-risk request in the chat: Force-push to main, with Deny and Hold to approve.
  • Codex asks which login method to add, with two numbered options and a typed answer below.

A real app, not a slideshow

People judge an experience by using it, not by looking at screenshots.

Open the prototype (opens in a new tab)

adnnankhan.github.io/Nudge/app

  • Anywhere, installablePhone, laptop or desktop, and it installs to the home screen.
  • Real interactionsHold to approve, Face ID, typing, voice and requests arriving on their own.
  • Built against the designsMade in Claude Code from detailed specs, tested in short rounds.

13Feedback and iteration

  1. New task before feedback: a form with Computer, Agent, Project folder, Instruction and Safety fields and a Start button.
    Before: New task form
  2. New task after feedback: Who should do it? with Claude Code, Codex and Cursor tiles, then recent tasks.
    After: pick an agent
  3. Then a normal chat: What should Claude Code do?, the last used project shown as one line, suggestions and the message box.
    Then a normal chat
  4. A Home card opened as its chat: the agent’s messages and the approval card inside the conversation.
    Card → its chat

Tested with people who use agents every day

The people who’d use Nudge are the ones who can tell what’s missing.

Chat, without becoming a chat app

Its job is still to get you the right decision fast.

Feedback

A chat-based way to work was missing

Chat in three places onlyA Home card opens its chat, New task opens a chat after you pick an agent, and Tasks lists every task. Home stays the approval hub.
Patterns people already knowLike Claude and ChatGPT (Jakob’s Law): your messages on the right, the agent’s as plain text, tool activity as one collapsible line.
Feedback

Adding a new task felt complicated

New task lost its formFilling fields was the friction people described. Your last project and computer are used automatically, shown as one line you can tap to change.
“Ask me”, not everythingThe agent comes to you only when you want it to: before every action, only before commands, or just when it’s done.

Nothing asks twice

I walked the flows as a frustrated user and removed every repeated question, because each one is a moment someone gives up.

  • New task remembers your choices
  • Cards don’t repeat who and where
  • Voice drops into the message box
  • An open app shows one banner
  • Home keeps your place
  • Low-risk requests approved together
  • One check covers low and medium for 5 min

Thanks for reading

Nudge makes sure people never miss what their coding agents need, and can answer correctly and safely from wherever they are.

Open the prototype (opens in a new tab) Say hello