Muskan Bhargava

Muskan Bhargava

Muskan Bhargava

Muskan Bhargava

Designing a Context-Aware AI Writing Assistant: that drove 2.1× higher email response rates

Designing a Context-Aware AI Writing Assistant: that drove 2.1× higher email response rates

Designing a Context-Aware AI Writing Assistant: that drove 2.1× higher email response rates

An assistant-style experience inside Gmail, helping recruiters and sales reps write more personalized, high-converting outreach using GPT.

An assistant-style experience inside Gmail, helping recruiters and sales reps write more personalized, high-converting outreach using GPT.

An assistant-style experience inside Gmail, helping recruiters and sales reps write more personalized, high-converting outreach using GPT.

Client

Client

ContactOut

ContactOut

Timeline

Timeline

Q3 2023

Q3 2023

Platform

Platform

Extension and Gmail

Extension and Gmail

Introduction 👈

Introduction 👈

Overview

ContactOut’s Chrome Extension is used by 1M+ recruiters and sales professionals to find verified contact details. Cold outreach is critical but often slow, repetitive, and hard to personalize.


Users were already experimenting with ChatGPT, but switching tabs, copying data, and pasting into Gmail was clunky and unscalable.


We saw an opportunity to embed a context-aware assistant directly in Gmail, reducing friction and improving outcomes.

Overview

ContactOut’s Chrome Extension is used by 1M+ recruiters and sales professionals to find verified contact details. Cold outreach is critical but often slow, repetitive, and hard to personalize.


Users were already experimenting with ChatGPT, but switching tabs, copying data, and pasting into Gmail was clunky and unscalable.


We saw an opportunity to embed a context-aware assistant directly in Gmail, reducing friction and improving outcomes.

Introduction 👈

Overview

ContactOut’s Chrome Extension is used by 1M+ recruiters and sales professionals to find verified contact details. Cold outreach is critical but often slow, repetitive, and hard to personalize.


Users were already experimenting with ChatGPT, but switching tabs, copying data, and pasting into Gmail was clunky and unscalable.


We saw an opportunity to embed a context-aware assistant directly in Gmail, reducing friction and improving outcomes.

The Opportunity

Our north star: help users follow up the moment they find a lead.
That meant:

  • Triggering email generation right after a profile was revealed

  • Embedding it into Gmail with minimal UI disruption

  • Personalizing content using LinkedIn data already available in our system. (Key differentiator in a market flooded with GPT-backed tools offering generic content generation)

The Opportunity

Our north star: help users follow up the moment they find a lead.
That meant:

  • Triggering email generation right after a profile was revealed

  • Embedding it into Gmail with minimal UI disruption

  • Personalizing content using LinkedIn data already available in our system. (Key differentiator in a market flooded with GPT-backed tools offering generic content generation)

Stakeholders

  • Product & Growth: I led UX, research, design, while focusing on activation, retention, and adoption loops

  • Engineering: Delivered complex DOM integration with Gmail and backend prompt handling

  • AI/ML Team: Defined token usage limits, prompt formatting, and OpenAI reliability handling

Stakeholders

  • Product & Growth: I led UX, research, design, while focusing on activation, retention, and adoption loops

  • Engineering: Delivered complex DOM integration with Gmail and backend prompt handling

  • AI/ML Team: Defined token usage limits, prompt formatting, and OpenAI reliability handling

My Role 👩🏻‍💻

My Role 👩🏻‍💻

Designed the assistant flow end-to-end inside Gmail

  • Created reusable Prompt library System

  • Defined prompt UX and personalization logic using LinkedIn profile data

  • Collaborated with ML on system prompts, token budgets, fallback design

  • Designed onboarding, awareness modals, and UI feedback states

  • Led user testing, analytics instrumentation, and post-launch iteration

Designed the assistant flow end-to-end inside Gmail

  • Created reusable Prompt library System

  • Defined prompt UX and personalization logic using LinkedIn profile data

  • Collaborated with ML on system prompts, token budgets, fallback design

  • Designed onboarding, awareness modals, and UI feedback states

  • Led user testing, analytics instrumentation, and post-launch iteration

The Opportunity

Our north star: help users follow up the moment they find a lead.
That meant:

  • Triggering email generation right after a profile was revealed

  • Embedding it into Gmail with minimal UI disruption

  • Personalizing content using LinkedIn data already available in our system. (Key differentiator in a market flooded with GPT-backed tools offering generic content generation)

Stakeholders

  • Product & Growth: I led UX, research, design, while focusing on activation, retention, and adoption loops

  • Engineering: Delivered complex DOM integration with Gmail and backend prompt handling

  • AI/ML Team: Defined token usage limits, prompt formatting, and OpenAI reliability handling

My Role 👩🏻‍💻

Designed the assistant flow end-to-end inside Gmail

  • Created reusable Prompt library System

  • Defined prompt UX and personalization logic using LinkedIn profile data

  • Collaborated with ML on system prompts, token budgets, fallback design

  • Designed onboarding, awareness modals, and UI feedback states

  • Led user testing, analytics instrumentation, and post-launch iteration

What success looked like for AI Email Writer ✨

What success looked like for AI Email Writer ✨

What success looked like for AI Email Writer ✨

We defined success not as GPT output, but how confidently and quickly a user could act. That meant minimal UI, and fast, tweakable outputs. The tool had to feel optional, but smart enough to become essential.


We defined a clear set of success metrics to guide MVP scope:

🎯 Prompt-to-send rate> 5% (achieved: 8.6%)

📥 Improve response rates by 2× vs non-personalized emails

⏱ Reduce time-to-send by 50% or more

📈 Drive adoption among Gmail users exposed to the assistant (+24%)


Timeframe: Phase 1 MVP shipped in 3 weeks, with follow-up personalization and library features in Phases 2–3.

We defined success not as GPT output, but how confidently and quickly a user could act. That meant minimal UI, and fast, tweakable outputs. The tool had to feel optional, but smart enough to become essential.


We defined a clear set of success metrics to guide MVP scope:

🎯 Prompt-to-send rate> 5% (achieved: 8.6%)

📥 Improve response rates by 2× vs non-personalized emails

⏱ Reduce time-to-send by 50% or more

📈 Drive adoption among Gmail users exposed to the assistant (+24%)


Timeframe: Phase 1 MVP shipped in 3 weeks, with follow-up personalization and library features in Phases 2–3.

We defined success not as GPT output, but how confidently and quickly a user could act. That meant minimal UI, and fast, tweakable outputs. The tool had to feel optional, but smart enough to become essential.


We defined a clear set of success metrics to guide MVP scope:

🎯 Prompt-to-send rate> 5% (achieved: 8.6%)

📥 Improve response rates by 2× vs non-personalized emails

⏱ Reduce time-to-send by 50% or more

📈 Drive adoption among Gmail users exposed to the assistant (+24%)


Timeframe: Phase 1 MVP shipped in 3 weeks, with follow-up personalization and library features in Phases 2–3.

What we thought users needed (and what the actually wanted) 🔍

What we thought users needed (and what the actually wanted) 🔍

What we thought users needed (and what the actually wanted) 🔍

Research Methods

  • 10 user interviews (recruiters + SDRs)

  • Internal survey across ContactOut recruiter and sales team

  • Competitive teardown of GrammarlyGO, Flowrite, Regie.ai

  • Prompt-response quality benchmarking using OpenAI

  • Onboarding modal A/B tests (GIF vs Video, CTA copy)

  • Product analytics (Metabase) + qualitative feedback from CS team


Key Insights

  • Users were most open to using AI when stuck in writing

  • Prompt-writing was a major skill gap, users didn’t know what to type

  • Most popular prompt use cases: follow-ups, recruiter intros, role-specific outreach

  • Contextual triggers (e.g., after “Reveal” or in reply thread) increased prompt usage

Research Methods

  • 10 user interviews (recruiters + SDRs)

  • Internal survey across ContactOut recruiter and sales team

  • Competitive teardown of GrammarlyGO, Flowrite, Regie.ai

  • Prompt-response quality benchmarking using OpenAI

  • Onboarding modal A/B tests (GIF vs Video, CTA copy)

  • Product analytics (Metabase) + qualitative feedback from CS team


Key Insights

  • Users were most open to using AI when stuck in writing

  • Prompt-writing was a major skill gap, users didn’t know what to type

  • Most popular prompt use cases: follow-ups, recruiter intros, role-specific outreach

  • Contextual triggers (e.g., after “Reveal” or in reply thread) increased prompt usage

Research Methods

  • 10 user interviews (recruiters + SDRs)

  • Internal survey across ContactOut recruiter and sales team

  • Competitive teardown of GrammarlyGO, Flowrite, Regie.ai

  • Prompt-response quality benchmarking using OpenAI

  • Onboarding modal A/B tests (GIF vs Video, CTA copy)

  • Product analytics (Metabase) + qualitative feedback from CS team


Key Insights

  • Users were most open to using AI when stuck in writing

  • Prompt-writing was a major skill gap, users didn’t know what to type

  • Most popular prompt use cases: follow-ups, recruiter intros, role-specific outreach

  • Contextual triggers (e.g., after “Reveal” or in reply thread) increased prompt usage

Assumptions validated or disproved

✅ Users wanted prefilled templates they could tweak

✅ Personalization based on LinkedIn profile increased reply rates

❌ Users would discover the AI feature naturally in the extension

❌ Chat-style UI preferred. Instead, users wanted minimal inline guidance

Assumptions validated or disproved

✅ Users wanted prefilled templates they could tweak

✅ Personalization based on LinkedIn profile increased reply rates

❌ Users would discover the AI feature naturally in the extension

❌ Chat-style UI preferred. Instead, users wanted minimal inline guidance

Assumptions validated or disproved

✅ Users wanted prefilled templates they could tweak

✅ Personalization based on LinkedIn profile increased reply rates

❌ Users would discover the AI feature naturally in the extension

❌ Chat-style UI preferred. Instead, users wanted minimal inline guidance

Design Process 🪄

Design Process 🪄

Design Process 🪄

🧪 Phase 1: MVP Prototype (Basic Prompt Field in Gmail)

We started with the simplest version: a single-line prompt input field at the top of Gmail compose


We shipped it fast.

Tested it with internal users + a small beta group.

Results

We found a key usability issue during interviews:

“I thought I was supposed to type the email itself here.”

Users were typing body text, not prompts.

🧪 Phase 1: MVP Prototype (Basic Prompt Field in Gmail)

We started with the simplest version: a single-line prompt input field at the top of Gmail compose


We shipped it fast.

Tested it with internal users + a small beta group.

Results

We found a key usability issue during interviews:

“I thought I was supposed to type the email itself here.”

Users were typing body text, not prompts.

🧪 Phase 1: MVP Prototype (Basic Prompt Field in Gmail)

We started with the simplest version: a single-line prompt input field at the top of Gmail compose


We shipped it fast.

Tested it with internal users + a small beta group.

Results

We found a key usability issue during interviews:

“I thought I was supposed to type the email itself here.”

Users were typing body text, not prompts.

🧪 Phase 2: Better Placement & Prompt Templates

We reworked the UI:

  • Moved input below the subject line, aligned with natural email flow

  • Introduced prompt placeholder examples

  • Added saved prompts dropdown

  • Tooltip help on hover


User testing:

  • Conducted 2 rounds of moderated testing (7 users each)

  • Showed 10+ UI prototypes across compose, reply, modal formats


Feedback improved significantly:

  • Users felt this version made more sense

  • Prompt usage rose +40% from previous version

  • But… editing prompts was still not seamless


From user interviews:

“I want to tweak the prompt without deleting everything.”
“Where do I save my own prompt?”

🧪 Phase 2: Better Placement & Prompt Templates

We reworked the UI:

  • Moved input below the subject line, aligned with natural email flow

  • Introduced prompt placeholder examples

  • Added saved prompts dropdown

  • Tooltip help on hover


User testing:

  • Conducted 2 rounds of moderated testing (7 users each)

  • Showed 10+ UI prototypes across compose, reply, modal formats


Feedback improved significantly:

  • Users felt this version made more sense

  • Prompt usage rose +40% from previous version

  • But… editing prompts was still not seamless


From user interviews:

“I want to tweak the prompt without deleting everything.”
“Where do I save my own prompt?”

🧪 Phase 2: Better Placement & Prompt Templates

We reworked the UI:

  • Moved input below the subject line, aligned with natural email flow

  • Introduced prompt placeholder examples

  • Added saved prompts dropdown

  • Tooltip help on hover


User testing:

  • Conducted 2 rounds of moderated testing (7 users each)

  • Showed 10+ UI prototypes across compose, reply, modal formats


Feedback improved significantly:

  • Users felt this version made more sense

  • Prompt usage rose +40% from previous version

  • But… editing prompts was still not seamless


From user interviews:

“I want to tweak the prompt without deleting everything.”
“Where do I save my own prompt?”

🧪 Phase 3: From Prompt Field to AI Assistant System

We upgraded the system to:

  • Expand input height (up to 5 lines with scroll)

  • Collapse after generation for less clutter

  • Add hover prompt guide tooltip ("Tips for mastering prompts")

  • Enable bookmarked prompts saved in Prompt Library

  • “Recreate” or “Refine” the output in 1 click

  • Added fallback state to show suggestion instead of error

We also introduced context scraping:

  • On replies, pulled in the last 5 emails in the thread

  • Prompt passed to backend with full thread context

  • GPT generated personalized responses that felt smart

Result:
✅ Average time-to-send dropped 63%
✅ Prompt reuse increased (1,200+ custom prompts saved in 30 days)
✅ Response rate doubled for emails using AI writer

🧪 Phase 3: From Prompt Field to AI Assistant System

We upgraded the system to:

  • Expand input height (up to 5 lines with scroll)

  • Collapse after generation for less clutter

  • Add hover prompt guide tooltip ("Tips for mastering prompts")

  • Enable bookmarked prompts saved in Prompt Library

  • “Recreate” or “Refine” the output in 1 click

  • Added fallback state to show suggestion instead of error

We also introduced context scraping:

  • On replies, pulled in the last 5 emails in the thread

  • Prompt passed to backend with full thread context

  • GPT generated personalized responses that felt smart

Result:
✅ Average time-to-send dropped 63%
✅ Prompt reuse increased (1,200+ custom prompts saved in 30 days)
✅ Response rate doubled for emails using AI writer

🧪 Phase 3: From Prompt Field to AI Assistant System

We upgraded the system to:

  • Expand input height (up to 5 lines with scroll)

  • Collapse after generation for less clutter

  • Add hover prompt guide tooltip ("Tips for mastering prompts")

  • Enable bookmarked prompts saved in Prompt Library

  • “Recreate” or “Refine” the output in 1 click

  • Added fallback state to show suggestion instead of error

We also introduced context scraping:

  • On replies, pulled in the last 5 emails in the thread

  • Prompt passed to backend with full thread context

  • GPT generated personalized responses that felt smart

Result:
✅ Average time-to-send dropped 63%
✅ Prompt reuse increased (1,200+ custom prompts saved in 30 days)
✅ Response rate doubled for emails using AI writer

📊 Final Iteration

  • Implemented AI Prompt Library (web + extension tabs)

  • Allowed users to save their best prompts

  • Dropdowns in Gmail + Extension both fetched saved prompts

  • Prompts now had system prompts per use case (Compose vs Reply)

  • Foundation used in other features: e.g., LinkedIn Commenter, Email Personalize

📊 Final Iteration

  • Implemented AI Prompt Library (web + extension tabs)

  • Allowed users to save their best prompts

  • Dropdowns in Gmail + Extension both fetched saved prompts

  • Prompts now had system prompts per use case (Compose vs Reply)

  • Foundation used in other features: e.g., LinkedIn Commenter, Email Personalize

📊 Final Iteration

  • Implemented AI Prompt Library (web + extension tabs)

  • Allowed users to save their best prompts

  • Dropdowns in Gmail + Extension both fetched saved prompts

  • Prompts now had system prompts per use case (Compose vs Reply)

  • Foundation used in other features: e.g., LinkedIn Commenter, Email Personalize

Awareness and Adoption 🪄

Awareness and Adoption 🪄

Awareness and Adoption 🪄

How did users discover the AI Email Writer?

We learned early that most users never saw the feature. Over a couple of releases, we experimented with different entry points to drive awareness.

How did users discover the AI Email Writer?

We learned early that most users never saw the feature. Over a couple of releases, we experimented with different entry points to drive awareness.

How did users discover the AI Email Writer?

We learned early that most users never saw the feature. Over a couple of releases, we experimented with different entry points to drive awareness.

We decided the sweet spot was a post-reveal prompt in the extension, supported by a one-time Gmail modal. This balance gave users contextual entry points without overwhelming them.

We decided the sweet spot was a post-reveal prompt in the extension, supported by a one-time Gmail modal. This balance gave users contextual entry points without overwhelming them.

We decided the sweet spot was a post-reveal prompt in the extension, supported by a one-time Gmail modal. This balance gave users contextual entry points without overwhelming them.

Key Insight! Iterative testing showed that just-in-time prompts worked best. Awareness rose +24%, and Day-7 retention doubled (5% → 11%).

Key Insight! Iterative testing showed that just-in-time prompts worked best. Awareness rose +24%, and Day-7 retention doubled (5% → 11%).

Key Insight! Iterative testing showed that just-in-time prompts worked best. Awareness rose +24%, and Day-7 retention doubled (5% → 11%).

Making GPT helpful (not overwhelming) 🤖

Making GPT helpful (not overwhelming) 🤖

Making GPT helpful (not overwhelming) 🤖

Working with our AI/ML team, we:

  • Used OpenAI GPT-3.5-turbo (later GPT-4 selectively)

  • Set a max token limit of 512 to keep output focused and lightweight

  • Created separate system prompts for Compose and Reply use cases

  • Prompt examples tested: 300+ across recruiter, sales, casual, formal tones

  • Prompt outputs logged and reviewed internally for tuning

Working with our AI/ML team, we:

  • Used OpenAI GPT-3.5-turbo (later GPT-4 selectively)

  • Set a max token limit of 512 to keep output focused and lightweight

  • Created separate system prompts for Compose and Reply use cases

  • Prompt examples tested: 300+ across recruiter, sales, casual, formal tones

  • Prompt outputs logged and reviewed internally for tuning

Working with our AI/ML team, we:

  • Used OpenAI GPT-3.5-turbo (later GPT-4 selectively)

  • Set a max token limit of 512 to keep output focused and lightweight

  • Created separate system prompts for Compose and Reply use cases

  • Prompt examples tested: 300+ across recruiter, sales, casual, formal tones

  • Prompt outputs logged and reviewed internally for tuning

Outcomes and Impact 📈

Outcomes and Impact 📈

Outcomes and Impact 📈

UX Impact

  • Reduced friction in email drafting at the moment of decision

  • Personalized prompts increased confidence and ease-of-use

  • Enabled faster workflows without leaving Gmail


Business Impact

  • GPT credit usage tied to monetization: 27% of all credit usage

  • Re-engaged dormant users through personalized outreach flow

  • Created long-term differentiator via profile-based AI personalization

  • Validated AI as a scalable revenue lever, paving the way for other AI assistant features


Design System Wins

  • Prompt Library reused in other features: LinkedIn Commenter, AI Email Personalizer

  • Onboarding UI (modal + CTA) became a blueprint for other AI features

  • Foundation for reusable Compose vs Reply system prompts

UX Impact

  • Reduced friction in email drafting at the moment of decision

  • Personalized prompts increased confidence and ease-of-use

  • Enabled faster workflows without leaving Gmail


Business Impact

  • GPT credit usage tied to monetization: 27% of all credit usage

  • Re-engaged dormant users through personalized outreach flow

  • Created long-term differentiator via profile-based AI personalization

  • Validated AI as a scalable revenue lever, paving the way for other AI assistant features


Design System Wins

  • Prompt Library reused in other features: LinkedIn Commenter, AI Email Personalizer

  • Onboarding UI (modal + CTA) became a blueprint for other AI features

  • Foundation for reusable Compose vs Reply system prompts

UX Impact

  • Reduced friction in email drafting at the moment of decision

  • Personalized prompts increased confidence and ease-of-use

  • Enabled faster workflows without leaving Gmail


Business Impact

  • GPT credit usage tied to monetization: 27% of all credit usage

  • Re-engaged dormant users through personalized outreach flow

  • Created long-term differentiator via profile-based AI personalization

  • Validated AI as a scalable revenue lever, paving the way for other AI assistant features


Design System Wins

  • Prompt Library reused in other features: LinkedIn Commenter, AI Email Personalizer

  • Onboarding UI (modal + CTA) became a blueprint for other AI features

  • Foundation for reusable Compose vs Reply system prompts

Learnings 💭

Learnings 💭

Learnings 💭

Building the AI Email Writer pushed me to think beyond interface design and into the deeper layers of human-assistant interaction, systems design, and user psychology.

  • Build for Confidence, Not Just Output
    A great assistant doesn’t just generate, it reassures, teaches, and guides. Helping users feel confident enough to hit "send" was the real success metric.

  • Start scrappy, but measure ruthlessly
    We shipped Phase 1 in under 3 weeks, minimal UI, maximal learning. Quick feedback loops helped us validate before investing more deeply.

  • Adoption ≠ Awareness
    Discovery touchpoints like modals, CTAs, post-reveal prompts played a key role in closing the loop between value and visibility.