EnterpriseSG research assistant

Key features

The beta findings shaped the features that followed. Here are some of the key ones, the problem each solved and the decisions behind them.

Skills and prompt enhancement

Beta feedback made it clear that officers weren’t struggling to access information. They struggled to ask for it in a way that produced useful results. Different officers expected different formats, tone and depth, and the agent had no way to account for that by default. I designed two features that tackle this at different points.

1. Skills

Skills are small, reusable instruction bundles that teach the agent how to execute a specific task consistently. For example, an officer activating the “Sector Landscape Summary” skill gets a structured overview of market size, key players and trends, without spelling any of that out in every prompt.

The challenge was making skills feel lightweight and within reach, not like a settings menu buried in a corner. The skills bar sits directly beneath the input, always visible. A manage link opens the library, where a drawer lets officers preview a skill while keeping the list in view.

2. Prompt enhancement

Even with skills active, officers sometimes sent prompts too brief or ambiguous to produce a good result. The Enhance button tackles this at the moment of input, expanding and clarifying a prompt before it’s sent, without the officer needing to know what a “good” prompt looks like.

I designed six states: empty (disabled), below the character threshold, ready, streaming (the text rewrites inline instead of showing a spinner), complete with two-level undo, and error, with amber for a timeout and red for a failure.

Company scoping

Officers needed a way to point queries at specific companies without breaking the flow of starting a conversation. Early versions treated scope as a single-company toggle, but feedback made it clear that scope had to support multiple companies.

Key constraint: scope had to be set before the first message. Changing it mid-session would invalidate the conversation context, so the entry point had to be clear, and the locked state had to explain itself rather than surface as an error.

Getting to the right pattern took several iterations, each exposing a different problem.

Iteration 1

Slash command

Typing “/” in the input opened a contextual menu with companies and skills, a pattern familiar from tools like Notion or Claude. The problem was discoverability: officers who didn’t know the shortcut would never find scope or skills at all.

Iteration 2

Settings gear popover

A single gear icon popover gave scope and skills one tidy entry point. But the icon gave no signal about what was configured: officers couldn’t tell whether scope was set or skills were active without clicking to check.

Iteration 3

Scope banner above input

A persistent “Scoped to” banner above the input made scope visible at a glance. But with scope above and skills below, the composer felt heavy: two rows of configuration before officers had typed anything.

Final

Config line

The solution was collapsing scope and skills into a single compact strip below the input: the config line. It always states the current setup (eg. “4 companies · 3 skills active”) and stays visible throughout the session, so officers can read their context at a glance before typing.

Clicking Edit opens a right-side drawer where scoped companies appear as chips, and officers can search to add or remove them. Changes are staged before they’re applied, so half-finished edits never take effect. Once confirmed, the config line and input placeholder update to match.

Source attribution

From the earliest interviews, trust was the central concern: officers wanted to know how a response was reached and where it came from before relying on it. So source attribution had to be visible at every stage of the research, not just appended to a finished answer. Two components carry this across the product.

1. Citation bubbles in reasoning steps

As the agent works through a query, it shows the web sources consulted at each reasoning step, inline and in real time, so officers see the research as it happens. Citation bubbles are compact chips on each step that expand to show source detail on demand.

2. Call report drawer

For internal sources, officers needed to verify the call reports a response cites without leaving the chat. Clicking a report reference opens a right-side drawer with the full report. The left sidebar collapses to an icon rail to make space, keeping the chat visible beside the drawer.

The key decision was treating the collapsed sidebar as a state triggered by the drawer, not a user preference. When the drawer closes, the sidebar expands again automatically, so officers never have to restore their workspace by hand.