Rebuilding Talview’s question library around skills

Experts searching Talview’s question bank for one skill found thirty near-identical questions and couldn’t tell which to trust. I rebuilt the library so every question lives under the skill it tests, and duplicates fell 95%.

Context

Talview makes hiring and assessment software, and every test it runs is built from a shared bank of questions. This was my first major project as an intern, and one of the highest-value problems in the product. I owned the design end to end and reported directly to the CTO.

One search, thirty answers

The design had aged, and questions were organised only by tags that everyone wrote differently. A search for “data analysis” returned thirty that looked almost the same. The library was hard to navigate, full of duplicates, and closed to AI, which had nothing structured to read.

$60K

of duplicate questions and cleanup, found in the content audit

Written first, filed later

You wrote a question first and decided where it belonged after. Tags were an afterthought: some went on as a question was written, some long after, and everyone tagged differently. That is how the copies kept coming. Shopify called it confusing, and nearly dropped the product over it.

Fifty ways to fix a table

My first idea was a better table: tidier columns, smarter filters, inline tags. With AI, I explored more than fifty directions. Every one was easier to scan, and every one was still a table.

He was right. I had been polishing the problem instead of solving it.

What makes a question worth trusting?

My PM told me to step back, so I asked why nobody trusted the questions. You trust one when you can see what it tests and the ones next to it. So I moved every question from tags to skill, the way Finder or Figma’s layers panel holds files. A structure people can read is one AI can read too.

Two builds before it held up

Once the model clicked, the wireframe took an evening. Surviving real content took two more builds. The first squeezed folders into the old table, and the second got the layout right but buried it in clutter. I rebuilt it before leadership ever saw it.

Walk to it, or search for it

Skills sit on the left like folders, so you can walk down to a question with the full path on top. Or search for it: every result shows the skill it belongs to, so the right one is easy to spot.

Judge it without opening it

Each skill lists its questions with the four things an expert checks before reusing one: its type, how many candidates have seen it, its status, and who changed it last. The everyday decisions happen right here, without opening anything.

Every change has a name

Open a question and its history comes with it: who changed it, when, and the exact words, with the review beside the change. A question is trusted because it can be traced.

Write it where it belongs

A new question starts inside the skill you’re looking at. Pick a type and the editor opens right there, filed before it has a title.

Three clicks, not three pages

Finding a question meant a search and three pages of lookalikes; now it is three clicks to the five in its skill. Writing one no longer starts with four fields, and the skill comes from the folder.

Duplicates fell 95%

Adoption rose, customers called the library much more intuitive, and Shopify stayed. Every question now carries its skill, type and difficulty, so our AI tags the ones nobody did and drafts new ones in the same shape.

95%

fewer duplicate questions

Structure compounds

Fifty better tables taught me that the table itself was the problem. AI made fifty directions cheap, and none of them asked what makes a question worth trusting. That question was the job. One decision, organising by skill, gave us finding, reuse and an AI that writes from the bank.

1. Every question under a skill
2. Found by walking to it
3. Reused, not rewritten
4. Written by AI from the bank

Next case study

Talview’s sign-in, from 75 seconds to 35