information-architecture
Produces the structure underneath the screens: what content exists, how it groups, what each group is called, and how people move between groups. The output is a sitemap someone can build navigation from, plus the test that would prove users can find things in it. Sits below user-journey-map (which says what people are trying to do) and above content-design (what one page says) and ux-writing (the exact strings). A structure nobody has tested is a hypothesis; this skill says so and ships the test.
How to respond
Get the inputs in one message. Ask at most 2 questions, spent on the weakest of: who the main audiences are and their top tasks, and what content exists today (a page list, current menu, URL export, or feature list). Default everything else: web, desktop and mobile, existing content kept unless obviously dead. With no content list, build the inventory from what the user described and mark it
[inferred].Inventory before structure. List every content item or feature as a flat table: item, type (task, reference, marketing, account, legal), audience, and a status (
keep/merge/cut/[inferred]). Grouping comes after the list, never before it. Why: starting from the old menu copies its mistakes into the new one.Group by user task and mental model, not by org chart. ✅ "Get paid", "Manage your team", "Taxes" — ❌ "Finance Ops", "People Platform", "Compliance Team". Pick one organising scheme per level and say which: task, audience, topic, or object. Mixing schemes at one level ("Products", "For developers", "Pricing", "Help") is the most common IA defect; when audience-based top-level nav is unavoidable, say why and flag the overlap risk.
Label in the user's words. Each group label is the term users would type into search or say aloud. Prefer short, concrete nouns or verb phrases; no internal product names unless users already know them; no "Solutions", "Resources" or "More" as a bucket. Record every label decision in the rationale table: label, what it holds, why this word, rejected alternatives, and the evidence (
[search logs],[support tickets],[interview],[assumption]).Split the navigation model into three systems:
- Global nav — 5–7 top-level items, present on every page. More than 7 means the grouping is not done.
- Local nav — the sections within one global area (sidebar, tabs, sub-menu).
- Utility nav — account, settings, help, search, language, sign out. Kept out of global nav. Also name cross-links (related content that lives elsewhere) and the search role: search backs up navigation, it does not replace it.
Keep depth and breadth honest. Default to 3 levels or fewer for anything users reach weekly. Put the top 3 tasks one click from the home or landing page. Flag every item that sits in two places (allowed for cross-links, not as two homes) and every group with one child (merge it up) or more than ~10 (split it).
Name the findability risks. For each risk: the item, why users will look in the wrong place, the severity (🟥 top task affected / 🟧 frequent task / 🟨 rare task), and the mitigation (cross-link, rename, move, search synonym). Ambiguous labels, overlapping groups and items that fit two parents are the usual sources.
Write the validation plan from
reference.md: an open or hybrid card sort when the grouping or labels are unproven, a tree test when the structure exists and needs proving. The tree test gets 6–10 tasks written as user goals without the label words in them (✅ "You were charged twice this month. Where would you go?" — ❌ "Find Billing"), each with the correct destination and a pass threshold. State participants and what result would change the structure.Emit with
templates/ia-spec.mdin one message: audiences and top tasks, inventory, organising scheme per level, the indented sitemap tree, the Mermaid diagram, the nav model, the label rationale table, findability risks, and the validation plan. Hand-offs: page-level content for a key page →content-design; menu and button strings →ux-writing; requirements for a new section →prd-draft. Seeexamples/payroll-app-nav.md.
Non-interactive runs (subagent, CI, headless): a missing content list becomes an [inferred] inventory and missing audiences become [assumption] lines at the top. If neither the product nor its content can be identified from the input, emit BLOCKED: need the product and its content or features and build nothing.
Useful references in this skill
reference.md— organising schemes, card sort and tree test methods, participant counts, metrics and pass thresholds, label teststemplates/ia-spec.md— the IA spec skeleton: inventory, tree, Mermaid, nav model, rationale and risk tables, validation planexamples/payroll-app-nav.md— worked example: a cluttered payroll app menu restructured around tasks
Quality bar
- The inventory exists and comes first. Every item in the sitemap traces to an inventory row; every
cutis listed, not silently dropped. - One organising scheme per level, named. Mixed schemes are flagged with the reason.
- Labels are user words with a recorded reason. Every global and local label appears in the rationale table with its evidence tag.
- Global nav has 5–7 items; utility items are not in it.
- Top tasks are one click from the landing page and depth is 3 levels or fewer for weekly tasks, or the exception is stated.
- The tree and the Mermaid diagram match node for node, and the Mermaid renders.
- Findability risks are named with severity and a mitigation.
- The validation plan has tasks written without label words, correct destinations, a threshold, and what result would change the design.
- Nothing is invented. Content items, audiences and evidence not in the input are tagged
[inferred]or[assumption].
When to use this skill
- ✅ A new product, site, app or help centre needs a sitemap and navigation
- ✅ "Users can't find anything" — restructuring an existing menu or content set
- ✅ Merging two products or sites into one navigation
- ✅ Deciding where a new feature or section should live
When NOT to use this skill
- ❌ The steps and emotions of one persona reaching an outcome — that's
user-journey-map - ❌ What one page should say and in what order — that's
content-design - ❌ The exact wording of menu items, buttons and errors across a flow — that's
ux-writing - ❌ The requirements and scope of a new feature — that's
prd-draft
Anti-patterns to avoid
- ❌ The org chart as navigation. Groups named after the teams that own them. Users don't know your teams.
- ❌ "Resources", "Solutions", "More". A bucket label means the grouping stopped before it finished.
- ❌ Redrawing the old menu. Skipping the inventory and rearranging the existing items keeps every old overlap.
- ❌ Two homes for one item. An item that lives in both "Settings" and "Team" will be maintained in one and stale in the other. One home, cross-links elsewhere.
- ❌ Search as the fix. "They can search for it" hides a structure problem; search logs full of nav-label queries are the symptom.
- ❌ Tree-test tasks that give away the answer. "Find the Invoices page" tests reading, not findability.
- ❌ Presenting an untested structure as final. Tag it a hypothesis and ship the test with it.