Singular Agent with Guided Prompts
Constrain each agent to one job and guide users through required inputs.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 98%
This interface pattern replaces an unrestricted chatbot with a narrowly scoped agent that produces one recognizable outcome. The designer first defines the job, then identifies every piece of information the agent needs and turns those requirements into a guided sequence of prompts. Users no longer have to infer capabilities or invent effective instructions from a blank text box. Constraints also make the system's limits visible, reducing the mistaken belief that a natural-language interface can do everything. The approach is especially useful in go-to-market functions, where a sales, support, or customer-success agent can gather structured inputs and deliver a specific artifact. Success depends on testing the guidance with non-technical users rather than assuming conversation alone is intuitive.
Origin
Extracted from Marketing Against The Grain, including Kieran Flanagan's Zapier onboarding experience and agent-design recommendation.
Core principles
- 01A text box without boundaries creates false expectations.
- 02One agent should own one clearly defined outcome.
- 03Guided questions reduce prompt-writing burden.
- 04Constrained interfaces make capability discoverable.
How to run it
- 1
Choose One Job
Define a single outcome the agent will reliably produce for a specific user and context.
Pro tip Name the output in terms users already understand.
Watch out Combining unrelated jobs recreates the ambiguity of a general chatbot.
- 2
Map Required Inputs
List the facts, files, decisions, and constraints needed to complete the job successfully.
Pro tip Separate mandatory inputs from optional enhancements.
Watch out Missing inputs will surface later as weak or fabricated output.
- 3
Build Guided Prompts
Ask users for the required information in a clear sequence, using examples or choices where useful.
Pro tip Ask one cognitively coherent question at a time.
Watch out A blank text box transfers the interface-design problem back to the user.
- 4
Constrain the Boundary
Redirect requests that fall outside the agent's defined job and explain the supported outcome.
Pro tip Offer a relevant next action instead of a dead-end refusal.
Watch out Pretending to support every request damages trust.
- 5
Deliver the Artifact
Produce the promised result in a consistent structure and expose any missing information or uncertainty.
Pro tip Make the output immediately usable in the target workflow.
Watch out Do not conceal incomplete inputs behind fluent prose.
- 6
Test with Target Users
Observe non-technical users completing the flow and revise prompts where they hesitate, misunderstand, or abandon it.
Pro tip Track both completion rate and downstream output quality.
Watch out Designer fluency with prompting is not representative of normal users.
In the wild
Instead of offering sales representatives a blank chatbot, a company creates an agent whose only job is producing an account brief. It asks for the company, target contact, sales objective, known objections, and relevant internal notes, then returns a standardized brief. Requests for unrelated content are redirected to the supported workflow.
→ Representatives complete the task without learning prompt engineering.
Common mistakes
Starting with a Blank Chatbox
Users may not know what the system can do or what information a good request requires.
Giving One Agent Many Jobs
Broad capability makes boundaries invisible and produces inconsistent interaction patterns.
Skipping User Observation
A flow that feels obvious to its builders may remain confusing to non-technical users.
Is it for you?
Best for
It is best for customer-facing agents that perform a repeatable sales, support, or success workflow.
Not ideal for
It is not ideal for expert users who genuinely require an open-ended general-purpose assistant.
From the transcript
“have a singular agent that does one thing and use a guided prompt.”
“the prompt says, Hey, give me this information, now give me that information, and here's the thing you need.”
“Rather than allow it to be a broad-based agent where you can do all kinds of different things, and the person kind of gets lost.”
From the episode
Ai Images: Flux, Spotting Fakes & 4 Marketing Use Cases