Gains:
- Ability to draft the menu, navigation and content structure (information architecture) with artificial intelligence and verify it with card sorting results
- Ability to draw user flows and screen-to-screen steps into artificial intelligence and capture logic gaps and edge cases
- Ability to critically adapt the structure suggested by artificial intelligence according to the real user language and mental model
If a product becomes difficult to use, the problem is often not with the color or the button, but with the structure. Information architecture (IA) is how content in a product is grouped, named, and arranged into a navigation structure: menus, categories, hierarchy. User flow is the sequence of screen-to-screen steps that a user goes through to achieve a specific goal (registering, purchasing a product). Both build the invisible skeleton. AI quickly produces an outline of this skeleton and helps catch logic gaps; But whether the structure fits the mind of the real user can only be determined by human judgment and testing.
Information architecture: speaking the user's language
The fundamental tension of IA is this: will you organize content according to the logic of the institution or the mind of the user? Most unsuccessful menus impose the company's internal department structure on the user ("Enterprise Solutions", "Operations Center"). If the user asks "How do I pay my bill?" he thinks.
Artificial intelligence is a two-way helper here. On the one hand, it suggests a reasonable initial category structure; On the other hand, it helps you test your current structure against the user's language. But the tags the model suggests are generally “average internet language”; The words your user actually uses may be different. Therefore, the IA draft is verified by card sorting: real users are given content cards and they sort them; you adjust the structure according to their grouping.
Tip: When getting category suggestions from artificial intelligence, also ask it to "translate each category into a question sentence that the user will construct." This speeds up the transition from corporate jargon to user language.
User flows: steps, decisions, and edge cases
A good flowchart makes three things clear: steps (what the user does), decision points (yes/no branches), and edge cases — situations off the main path: empty cart, bad password, internet outage, expired session. Inexperienced designs draw the “happy path” (the scenario where everything goes well) and forget about edge cases; The user mostly hangs out in these forgotten places.
This is where AI is powerful: when you export a stream, you ask "what edge states are missing?" When you ask, it usually catches a few you missed. The model works like a flow critic. But you decide whether each branch it recommends is really necessary, product scope and user reality; Otherwise, the flow becomes unnecessarily complicated.
stream element
Contribution of artificial intelligence
human decision
Main steps
Standard ranking draft
Product-specific simplification
decision points
Possible branches
What is really necessary
Edge cases
Reminder of forgotten things
What will be covered
Label/text
first suggestion
Adaptation to user language
Error paths
General patterns
Actual error types of the product
three mini cases
Case 1 — The menu has become simpler. A team gave 9 main menu titles to artificial intelligence and asked them to "translate them into the user's question sentence." It was seen that 3 headings corresponded to the same question ("How do I manage my account?"); Menu reduced from 9 to 6. Then 14 users were tested with card grouping and the merge was verified. Click success increased from 61% to 86%.
Case 2 — Caught edge case. In one recording run, the team had drawn only the successful recording. AI asks “what happens if the email is already registered?”, “if the verification email does not arrive?”, “what if it closes and returns midway?” generated questions. The team also added 3 edge states to the flow; When it went live, support requests decreased significantly. Lesson: the model is a good "reminder of the forgotten."
Case 3 — Redundant branch was rejected. AI suggested branches to a payment flow such as “cryptocurrency option” and “multi-address management.” These were not included in the scope of the product; the designer rejected the suggestions. Lesson: not every suggestion of the model is accepted; The scope decision is human's, otherwise the flow becomes bloated.
Copiable prompts
Your role: information architecture specialist. Group the following content/feature list into categories that suit the user's mind. Give each category both a short label and a question phrase that the user will construct. Avoid corporate jargon. Use up to 6 parent categories.List: <<contents>>
Examine this user flow: <<steps>>.1) List missing edge cases (idle state, error, timeout, re-login).2) Flag steps that seem unnecessary or out of scope.3) Suggest tags that do not fit the user language.Give each item a one-sentence justification. Strictly speaking; Say "must be checked".
Sketch a screen-by-screen user flow for the following task: "<<task>>".Format: each step = Screen name > user action > system response.Separate decision points with "IF... THEN". Show at least 2 error paths along with the happy path.
Compare this menu structure with the card grouping result: Recommended structure: <<a>> User grouping: <<b>>. Show incompatible placements in a table and justify which one is closer to the user's mind.
Weak prompt / Strong prompt
Weak: "Write me a recording stream."
Result: A generic step list with no edge cases, containing only the happy path.
Strong: "Draw a screen-by-screen flow for this task; give each step in the format 'screen > action > system response'; show at least 2 error paths next to the happy path; separate decision points with IF-THEN."
The result: a workable flow with visible edge states and branches.
Difference: requires strong prompt format + error path imperative + decision point; these bring the flow closer to the real world.
Common mistakes
- Just designing the happy way. The user gets stuck most often in edge situations; skipping them is the most expensive mistake.
- Bringing corporate jargon to the menu. The user searches for “where is my order,” not “operations center.”
- Finalizing IA without card grouping. Model proposal is hypothesis; user confirms the test.
- Adding each flow suggestion of the model. Out-of-scope branches swell the flow; Rejecting is also a decision.
- Not testing labels at all. A label that you think is "understandable" may be unfamiliar to the user.
In summary
Information architecture and user flows are the invisible skeleton of the product; The fluidity of the user experience largely comes from this. AI suggests plausible category structures, critiques flows, and is especially powerful at catching forgotten edge cases. But whether the labels match the user language is determined by card grouping, and the scope of the flow and the necessity of branching are determined by human judgment. Use the model as a drafter and critic; You calibrate the structure according to the user's mind.
Application task
- List 8-10 contents/features of a product and create a category structure with the first prompt.
- Translate each category into the user's question sentence and combine those that are confused.
- Select a task and draw a screen-by-screen flow with the third prompt (with at least 2 error paths).
- With the second prompt, criticize the flow, add missing edge cases, and remove unnecessary steps.
- If possible, group mini cards with 3-5 people and adjust the structure according to the results.
checklist
- [ ] I translated the categories into user language and removed the jargon.
- [ ] In addition to the happy path, I also drew error paths in the flow.
- [ ] I scanned the missing edge cases with artificial intelligence and added them.
- [ ] I consciously rejected out-of-scope suggestions.
- [ ] I tested the IA draft with card grouping or user testing.
- [ ] I verified the tags with real user words.