Go back

Case study / Concept project

Comic Buff: conversation design for a hobby nobody explains well

A chatbot that teaches new collectors to grade, value and authenticate comics, one question at a time. The content model came before the conversation, and that order is what made it answer in context.

25+intents across grading, glossary, trends and authenticity
4changes to the prototype, from testing with 3 collectors
Role
Content strategy, conversation design, prototyping, testing
Team
Solo
Timeline
One semester
Context
Concept project
Comic Buff: the persona the bot was designed for, a new collector who has always read comics

Comic Buff is a chatbot that teaches new collectors to grade, value and authenticate comics, one question at a time. I designed the content model before the conversation, and that order is what made the bot answer in context instead of by keyword. Testing with three collectors surfaced four problems, and each one changed the prototype.

In short

Problem
Comic-collecting knowledge lives in forums and dense articles that assume you already know the vocabulary. New collectors have to learn how to learn before they can start.
Opportunity
A conversation can teach in the order a beginner needs, answering one question and pointing to the next, which a forum thread or glossary can’t.
What I did
Audited the domain content, turned it into a taxonomy of intents, and designed a Voiceflow chatbot that teaches through short exchanges and suggests what to ask next.
Outcome
A working bot with 25+ intents covering grading, glossary, trends and authenticity, plus beginner and advanced learning paths. Testing with 3 collectors led to four changes.
Proudest decision
Treating the content audit as the design work. The taxonomy, not the chat UI, decides whether a bot answers in context.

01 / Context and constraints

A visual subject, in a text medium.

I designed for a 25-year-old graphic designer who has always read comics and now wants to collect them. He doesn’t know how grading works, can’t find a good place to start, and can’t judge value or authenticity.

  • Scripted, not generative. Voiceflow intents and flows meant every answer was authored and every question had to map to an intent.
  • A visual subject in a text medium. Grading is judged by eye: spine stress, corner wear, colour.
  • Short attention. People arrive with one question, not a syllabus.

02 / What research told me

Confidence dips at one question.

I mapped the persona’s journey across four stages: discovering the bot, deciding to try it, learning, and acting by walking into a comic store. His confidence dips sharply at one moment: when he asks “how do I actually start?” and gets back a wall of factors.

User journey map across awareness, decision, engagement and action, showing tasks, goals, touchpoints and a mood line that dips while learning to grade and recovers once he starts collecting.
The mood line dips at one question: how do I actually start?
The bot didn’t need more knowledge. It needed to hand that knowledge over in an order a beginner could follow.

03 / Key design decisions

Three decisions that shaped the bot.

01

Build the content model before the conversation

Context
Chatbots built flow-first tend to match keywords, not meaning.
Options
  • A — write flows for the obvious questions and patch gaps later
  • B — audit the domain first, then derive intents and flows from it
Choice
B. I built a content inventory from CGC grading standards, a glossary and collector blogs, then structured it into intents, flows and utterances.
Why
The taxonomy sets the boundaries of what the bot can answer, so replies land in context.
Trade-off
A slower first demo. Nothing talks until the model is done.
02

Lead the user instead of waiting for them

Context
Beginners don’t know what to ask. The empty prompt is where the journey-map dip happens.
Options
  • A — open text input that answers whatever is asked
  • B — a fixed menu
  • C — answer the question, then suggest two or three next steps along a learning path
Choice
C. Every answer ends with suggested next questions as tappable buttons, and “knowledge journeys” offer step-by-step beginner and advanced routes. Typing a question still works at any point.
Why
It keeps the freedom to ask anything while removing the blank page.
Trade-off
More flows to write and maintain, because every answer needs good next steps.
03

Define terms in passing, not in an appendix

Context
Jargon is the barrier, and a separate glossary means leaving the conversation to understand it.
Options
  • A — a glossary module users visit
  • B — define each term inline the first time it appears, with the glossary as a fallback
Choice
B, with the glossary kept as its own module for people who want to browse.
Why
Learning happens at the moment a term is needed.
Trade-off
Longer replies, which testing later flagged.

04 / Testing and what changed

Three collectors, four changes.

I tested the Voiceflow prototype with 3 collectors, in sessions averaging about 10 minutes. The sample was small, but all three sessions raised the same four issues.

FindingWhat it told meWhat I changed
Users wanted images alongside grading explanationsGrading is visual, and words alone don’t carry itAdded example images to the grading answers, showing what each kind of wear looks like
Retracing steps was hardUsers need a way out at every pointAdded persistent “Go back” and “Main menu” options to every reply
Users expected guidance, not a bot waiting for inputSuggestions had to be stronger and always presentEvery reply now ends with two or three suggested next questions
Some replies ran long, and attention droppedMy inline definitions made answers too denseSplit long answers into short messages, with “Tell me more” for the detail

05 / Outcomes

What the project produced.

25+intents across grading, glossary, trends and authenticity
5topics in the taxonomy the bot answers from
3collectors tested, in sessions of about 10 minutes
4changes the testing made to the prototype
  • A working Voiceflow bot with 25+ intents and dynamic topic branching.
  • A content inventory and taxonomy that serve as the bot’s source of truth.
  • Modular beginner and advanced learning paths that adapt to the user’s choices.
  • Lightweight desktop and mobile UI samples for a future web version.

06 / What I would do differently

What I would do differently.

Rebuild it on an LLM, grounded in the inventory

I’d keep the content inventory as the source the model answers from, so replies stay tied to CGC standards rather than the model’s guesses. High-stakes answers, like authenticity checks, would stay scripted, because a confident wrong answer costs a collector money. The 25 intents would become a test set.

Design for images from day one

Testing confirmed what the domain already said: grading is visual.

Set a reply-length budget before writing,

instead of trimming after testing.

Test with complete beginners

My testers were already collectors, so they knew more vocabulary than the persona.