NexusGPT
Enhancing first-crack AI agent usability
with guided prompt design
Timeline
2 months (Dec '24 - Jan '25)
My role
Product designer
Team
1 PD, 1 PM, 1 DEV
Status
Under-development
Background
1/7
Problem
3/7
AI agents aren’t finished when they’re deployed.
Just like human employees, they grow through feedback.
As they interact with real users, they encounter edge cases, learn new tasks, and develop blind spots.
But improving an agent isn’t as simple as checking a box. It means updating its brain — the prompt.
Problem
3/7
Problem
3/7
In theory, Nexus supports this loop—but in practice, the system makes this incredibly difficult.
But don't just take my word for it. Here's Justine, a customer support manager, to share what her experience was really like.
Problem
3/7
Justine's experience
Agent is deployed
Week 1
Issues start cropping up
Week 2-3
Fixes attempted
Week 4
Real consequences
Week 5
The agent works well and receives positive feedback.
Confusing responses. Edge cases encountered.
Unnecessary escalations to human support.
Repetitive or robotic tone.
Justine edits the agent’s prompt.
She’s overwhelmed by a wall of text.
She makes a hesitant change.
The agent’s logic breaks the next day.
Justine loses trust and stops editing.
😄
🤔
😥
💀
Justine's feelings throughout
Justine wasn't alone. Countless managers faced this same uncertainty every day.
The gap
Justine's story is far from unique. It exposes the fundamental flaws in how Nexus handles prompt maintenance, making agent improvement unnecessarily complex and risky.
Problem
2/7
Problem
3/7
We studied how users maintained agents and where their understanding fell short.
Problem
3/7
Lack of prompt expertise
Most managers don’t understand prompt engineering, leading to guesswork and ineffective updates.
Feedback isn't actionable
Feedback from conversations and tasks isn’t mapped to the right parts of the prompt, forcing users to hunt for issues blindly.
Fear of breaking things
Even small edits can cause unexpected failures, making users hesitant to improve their agents.
Problem
3/7
Watching how managers maintained their agents made one thing obvious: improving agents in Nexus often feels like pure guesswork and frustration.
Problem
3/7
Problem
3/7
Unstructured, monolithic prompts
Low prompt literacy
Feedback disconnected from prompts
Risky edits
Problem
3/7
All this added friction increased the burden on the dev team and left users engaging with a feature that failed to solve the core problem
Problem
3/7
>90%
created agents were undeployable
An overwhelming majority of agents created through the form were essentially useless.
80%
of most prompts needed edits
The generated prompts needed heavy edits, showing the form wasn’t doing its job.
6
hours lost weekly
The dev team manually fixed agent prompts through client calls, which drained limited dev resources.
Insights
4/7
Problem
3/7
After talking to devs and watching how they manually improved prompts, it became clear: our flow asked the wrong things — or didn’t ask enough.
Problem
3/7
Focused on “what,” not “how”
Agent creation = hiring. But our flow missed the basics: job, personality, training, onboarding.
Training came after the fact
Knowledge wasn't part of setup. Agents launched unprepared and underinformed.
Key fields were optional or too basic
Only name and job title were required. Purpose and capabilities? Skipped or vague.
No preview, no feedback
Users couldn’t see the prompt or agent evolve. No chance to course-correct early.
No real personality building
Just one “writing style” input. No deeper traits, tone, or values — agents felt flat.
Lack of immersion = lack of ownership
It felt like “filling out a form,” not “creating an agent.”
Problem
3/7
And on top of all that, the UI didn’t help — it was cramped, flat, and lacked any real visual hierarchy. Nothing about it felt like you were building something meaningful.
Research
5/7
Problem
3/7
This wasn't a form problem. It was a framing problem.
To ground the idea, I studied how real teams onboard roles like the ones users gave their agents — support reps, consultants & sales associates.
FINDINGS
CONCLUSION
The original agent form only covered ~20% of a typical onboarding process.
It captured surface-level inputs — job title, responsibilities, writing style — but missed the deeper layers: tone, behaviour, tools, training, and context. Most agents weren’t usable on the first try.
We didn’t need a better form.
We needed a better frame — one that treats agent creation like onboarding a teammate.
Problem
3/7
Rebuilding the flow around the onboarding model
We modeled the flow after real-life hiring to ask better, more insightful questions — helping users define agents more clearly and end up with stronger, more usable first drafts.
Rebuilding the flow around the onboarding model
Problem
3/7
Designs
6/7
Problem
3/7
I turned previously optional identity fields into a required first step — helping users define who the agent is from the start, reducing ambiguity for both the user and the system.
This step mirrors how real teams start onboarding — by clarifying who the person is, what they do, and where they belong. It’s about grounding the agent in a role and context before anything else.
Problem
3/7
Info that was optional before is now front and center.
A live preview gives the agent a face — making the process feel more personal and real.
Establishing a solid identity nudges users to think of the agent as a clearly defined persona with intent — not just a text generator.
Lets users pick from leading LLMs depending on the task — whether it needs speed, reasoning, creativity, or long context.
Problem
3/7
Prompts were often vague about the agent’s role, so we introduced agent types to give structure — and dramatically improve prompt quality.
Just like a real role needs a job description, agents need purpose. Users also chose what type of work was to be done. This raised prompt quality immediately, because now Nexus knows what to expect.
Problem
3/7
Problem
3/7
Selection pushes users to be specific about intent
Choosing a type helps structure the prompt behind the scenes
Problem
3/7
To prevent hallucinations and silence, users can now add knowledge and fallback logic during setup.
That’s why I designed the Knowledge step — to define not just what the agent should know, but how to retrieve it and what to do when it doesn’t.
Problem
3/7
Add knowledge when it matters — skip when it doesn’t
Define how the agent handles gaps or uncertainty
Reduce hallucination by grounding what the agent knows
Problem
3/7
I reduced deployment friction by enabling users to equip agents with the right tools and skills during creation — so they’re functionally ready from creation
Previously, agents were created first, and only after they failed at certain tasks did users realize they needed to attach skills (like search, API access, Airtable, etc.). This made agents feel limited or broken by default — and it often wasn’t obvious what was missing until something didn’t work.
Problem
3/7
Reduce launch friction by ensuring the agent is functionally ready
Relevant skill suggestions remove the guesswork
Problem
3/7
To address agents lacking personality or brand alignment, I improved the personality step with better questions and a more comprehensive tone builder.
Users can upload brand copy → we analyze tone → generate a matching voice. Or just pick a preset. Either way, agents sound way more on-brand.
Problem
3/7
Problem
3/7
To emphasize intentionality and immersion, I designed a live preview that updates in real time as users build their agent
We made the flow full-screen to feel more immersive. Inputs on the left, live agent preview on the right — so users can watch their agent come to life in real time. The snappy transitions are pretty cool too.
Problem
3/7
Problem
3/7
Conclusion
7/7
Problem
3/7
Still in dev, but early feedback is strong
No hard data yet, but users were way more confident during walkthroughs. We’ll measure prompt quality pre/post-launch to see how much this actually helps.
Rebuilding the flow around the onboarding model
Problem
3/7
Final thought: users don’t need power. they need clarity
We thought we were simplifying prompting. But the real win came from designing an experience that felt natural — something people already knew how to do.
Rebuilding the flow around the onboarding model
Return home
Back to work