← Back

Information Architecture Before Interface.

Bad information architecture has a distinctive sensation: you feel stupid in a product you otherwise respect.

You click the obvious label. It doesn’t go where it should. You back out. You try the second obvious label. Same problem. The interface is telling you, politely, that you don’t understand it.

And then you do what everyone does when a system humiliates them: you stop trusting it.

IA is the first interface

Navigation is the visible symptom. IA is the underlying structure. If the structure is wrong, the pixels are an expensive lie.

This is where prompting becomes genuinely powerful—not as idea generation, but as a demolition tool. AI is very good at two things humans are strangely sentimental about:

  1. killing synonyms
  2. being relentlessly literal about what words imply

A few tactics that save weeks

  • Write the task as a verb. “Cancel subscription,” “export data,” “recover account.” If you can’t verb it, you don’t understand it.
  • Run a synonym purge. Pick one verb per action. Delete/remove/archive are not triplets. They’re indecision.
  • Ban low-scent labels. “Resources.” “Solutions.” “Learn more.” These are hallway signs that point to more hallways.
  • Choose one primary mental model per section. By object (Files), by job (Billing), by audience (Teams). Mixing models is how navigation becomes a junk drawer.
  • Stress-test with edge cases. “Where would a tired person look for this at 2am?” If the answer is “it depends,” your labels are lying.
  • Use polyhierarchy sparingly. Cross-list only when ambiguity is real, not because you couldn’t decide.
  • Make ‘place’ obvious. Breadcrumbs, section headers, local nav. Users don’t just need routes—they need orientation.

Two prompts that improve IA fast

1) The label autopsy (keep / rename / delete).

You are an information architect. Be ruthless about clarity and information scent.
 
Input:
- Navigation tree (categories + labels)
- Target users and top tasks
 
Output:
For each label: KEEP / RENAME / DELETE and a one-sentence reason.
 
Rules:
- Ban vague labels (“resources,” “learn more,” “solutions,” “misc”).
- Prefer verb-first labels for actions.
- Enforce one verb per action across the product.
- If two labels overlap, propose a merge and name the survivor.

2) The tree-test simulator (where would users click?).

Given this navigation tree and these tasks, predict the first click a user would make.
 
Output a table:
Task | Likely first click | Confidence (high/med/low) | Why | Proposed fix (if low)
 
Constraints:
- Max 10 tasks.
- Fixes must be renames or regrouping, not a full redesign.

The point

IA work is not glamorous because it doesn’t produce a screenshot worthy of applause. It produces something rarer: a system that stops asking users to be mind readers.

If the names are wrong, the interface is already lying.