September 2025
Search shouldn’t feel like work.
Rethinking how context flows between conversational AI and traditional search results.
Problem
Great results.
No way to use them.
In agentic search, keyword results appeared alongside the AI response, but they were effectively static. To do anything with them, you had to refer to them in a prompt instead of interacting with them directly.
I found 3 records related to Acme Corp’s Q4 activity. The expansion opportunity is currently in negotiation, and there’s a recent forecast document attached.
Acme Corp — Q4 Expansion
OpportunityStage Negotiation
Q4 Revenue Forecast.pdf
FileShared Oct 2025
Project Atlas — Overview
KnowledgeUpdated Nov 2025
Ask a follow-up...
Prompt Overload
You had to spell it out. Every time.
There was no way to specify which result you were referring to, so you had to name it in the prompt. Referencing multiple results meant manually typing them out one by one, even when they were already on screen.
I found 3 records related to Acme Corp’s Q4 activity. The expansion opportunity is currently in negotiation, and there’s a recent forecast document attached.
Acme Corp — Q4 Expansion
OpportunityStage Negotiation
Q4 Revenue Forecast.pdf
FileShared Oct 2025
Project Atlas — Overview
KnowledgeUpdated Nov 2025
Ask a follow-up...
The interface had the context.
The interaction ignored it.
What if results weren’t just visible, but actionable?
I focused on turning search results into usable context for AI with clicks — not keystrokes.
One-Click AI Actions
One click. Done.
AI-powered actions are embedded directly on each result, using the record itself as context. What would normally require writing out a detailed prompt is reduced to a single click, with the agent already scoped to the right data.
I found 1 qualified lead and 1 open deal. Morgan Ellis at Acme Corp is the warmest lead — VP Sales, qualified this week.
Morgan Ellis — Acme Corp
LeadStatus Qualified
Acme Corp — Q4 Expansion
OpportunityStage Negotiation
Ask a follow-up...
Drag-to-Context
Bring results into the question. Literally.
Drag and drop turns results into context. Multiple records can be pulled directly into the chat composer, where they become part of the follow-up message without needing to be referenced or described explicitly.
I found 1 qualified lead and 1 open deal. Morgan Ellis at Acme Corp is the warmest lead — VP Sales, qualified this week.
Acme Corp — Q4 Expansion
OpportunityStage Negotiation
Morgan Ellis — Acme Corp
LeadStatus Qualified
Ask a follow-up...
Context persistence
Carry them through the conversation.
As results give way to the conversation, acted-on records remain visible in-thread as context. This preserves continuity through the transition, keeping the interaction tied to what was selected.
I found 1 qualified lead and 1 open deal. Morgan Ellis at Acme Corp is the warmest lead — VP Sales, qualified this week.
Morgan Ellis — Acme Corp
Lead
Morgan Ellis is VP of Sales at Acme Corp and was qualified earlier this week. She's tied to the $240K Q4 Expansion deal currently in negotiation, with legal reviewing redlines and a December 15 target close. For the call, lead with timeline urgency, confirm procurement requirements, and gauge whether executive alignment is needed to move forward.
Ask a follow-up...
Under the Hood
Simple on top.
Structured underneath.
Both interactions rely on the same underlying framework. The interaction determines how intent is captured, while reusable templates combine that intent with record context before sending a structured request to the agent.
Instructional prompt
Common instructions define how the agent interprets structured tasks.
Task template
Reusable templates combine user intent with selected record context.
UI metadata
Rendering metadata keeps the conversation readable without exposing prompts.
One-click actions
A predefined task template is selected and hydrated with the chosen record's details before being sent to the agent.
Chat rendition
Morgan Ellis — Acme Corp
Lead
Task template
Task:
Prepare the user for a sales call.
Return talking points, risks, and next steps.
Inputs:
objectType = {{objectType}}
recordId = {{recordId}}
Contextual follow-ups
The user-authored message is combined with selected records through a task template before being sent to the agent.
Chat rendition
Acme Corp — Q4 Expansion
Opportunity
Task template
Task:
Respond to the user's message.
Use the provided records as context.
Inputs:
userMessage = {{userMessage}}
records = {{records}}
From idea → prototype → roadmap.
This work explored new ways to blend traditional keyword search with AI capabilities without predefined product requirements or designs. The resulting concepts helped shape internal discussions around context-aware agentic experiences and informed future product direction.