Bot Instructions
Control what your bot says, how it responds and what it avoids.
Company
Zendesk AI Agents, Berlin
Role
Sr. Product Designer
Disclaimer: The designs, concepts, and product details shared here are intended solely to demonstrate my design process and contribution. All intellectual property remains with the respective companies. Confidential information has been omitted and certain details have been generalised where necessary.

Zendesk's zGPT could answer questions from knowledge sources, but gave users very little control over how the bot actually behaved. Four tones of voice, two of which were functionally identical, and two answer lengths. Bot Instructions changed that, letting customers write free-text instructions in natural language that could be added, edited, activated, or deleted at any time.
Problem statement
Businesses had no real way to control their own bot. To get it to say the right thing, teams were building workarounds, writing scripted conversations even when the answer was already in their help articles. It took a lot of time and nobody was happy with it. Competitors already had this feature. We lost a customer and a few potential deals because we didn’t, and even customers who signed with us asked for it to be a priority.
So the question was: how do we let businesses control their bot in their own words, without making it complicated or risky?
Who is this for?
Pretty much any business that wanted more than the preset options:
-
Brands with a distinct voice they wanted the bot to match
-
Teams with a style guide, with specific words, terms or formatting they always use
-
Teams tired of certain phrases, like too many apologies, “according to the sources…” or always pushing people to “contact support”
-
Teams with specific rules for how the bot should handle certain topics
Smaller businesses wanted something quick to set up. Larger ones wanted the bot to sound the same across every channel.
Understanding requirements
The requirements came straight from our CS team, who talked to customers every day, and from feedback by customers across streaming, fintech, e-commerce and retail.
One thing was clear: people wanted to talk to their bot the way they’d brief a new team member, in normal language and not through complicated settings.
I mapped out the main flow: adding an instruction, saving it, switching it on and editing it later. For each step I noted what the user does, what the system does and what we still didn’t know. This became the shared reference for the whole team.

.png)

What we had to get right
Letting people write their own instructions for an AI is harder than it sounds. We kept coming back to three worries:
-
Will the AI actually listen? If it ignores sensible instructions, the feature is useless. We tested early versions by throwing lots of instructions at it at once, including ones that contradicted each other.
-
Will people understand what it can and can’t do? Instructions change how the bot talks and behaves. They can’t make it know things that aren’t in the help articles. We didn’t want to overpromise.
-
What happens when people get it wrong? Without a safe way to try things out, one bad instruction could break a live bot and land on the CS team’s desk.
Each of these shaped the design below.
Design Solution
Simplified Layout
All instructions sit on one page, so anyone can see at a glance what’s shaping their bot, no digging through menus. Each instruction has a short name, the instruction itself, and a clear on/off status, so you can pause one without deleting it.
“Save” and “Save and activate” are separate buttons, so writing a draft never accidentally changes the live bot. You can also update many instructions at once.

Sandbox testing
Users can try out their instructions on their real bot before publishing, without customers seeing anything. This was the answer to our biggest worry: people breaking their bots and needing us to fix them.

Bulk testing
You can also pick several instructions and test them together right from the main page. That way you can see how they work with each other and spot clashes before anything goes live.

Releasing in stages
The easiest way for this project to fail was to keep adding “just one more thing” and never ship. So we kept each release small and got it to customers early:
-
Basic implementation: add, edit, delete and switch instructions on and off. Customers could try it and give us feedback early.
-
Conversation Logs: conversation logs show which instructions were active. Customers could test on a demo bot, then move to their live bot.
-
Beta: testing on the live bot without affecting real customers.
-
Later: instructions that only apply in certain cases (like a specific language or region), warnings when instructions clash, help writing better instructions, and version history.
Each of these shaped the design below.
The impact
This feature was tied directly to one of Zendesk’s biggest goals. When businesses can tell their bot exactly how to behave, it gives better answers. Better answers mean more customer questions get solved without a person needing to step in.
Increased automated resolution rate from
75%
of the instructions customers set up made the bot’s answers better
4.1%
to
9.4%
What I learnt
Designing for AI means designing for things you can’t fully predict. The design has to be honest about what the feature can do, show people what each change actually does, and make mistakes easy to undo. Testing and conversation logs weren’t extras. They were what made people trust it.
© 2026 by Rajashri Achar. Created on Wix Studio.