Hugh Williams' Claude Code search engine has an old-school secret

Why it matters

Zettair is a better signal than a toy demo: it shows coding agents can amplify deep domain expertise, but they do not remove the need for it.

Hugh Williams, a former Google engineering leader, used Anthropic's Claude Code to build Zettair, a working search engine that indexes 1.5 million Wikipedia articles, Business Insider reported Monday.

The headline version is simple: Williams says he wrote "zero lines of code" himself. The useful version is more interesting. Zettair is not a search engine conjured from a blank prompt by a nontechnical user. Business Insider reports that the underlying search engine is based on an information-retrieval system Williams helped build in the early 2000s. That makes the experiment a cleaner read on where coding agents are actually strong today: not replacing judgment, but giving a domain expert a faster route from old architecture to usable product.

Williams is not a random prompt engineer testing a toy app. Past roles include senior positions at Pivotal, eBay, and Microsoft Bing, per his 2014 UC Berkeley DataEDGE speaker bio. In April 2026, Business Insider reported that Paramount hired Williams for ad-tech product work, after earlier senior roles at Google and eBay.

That background matters because Zettair's feature list reads like a compressed history of search product expectations: autosuggest, query-biased snippets, related searches, trending topics and AI-generated summaries, according to Business Insider. The project is less a proof that search is suddenly easy than a proof that an engineer who already knows what matters in search can ask an agent to assemble the interface, glue and surrounding product surface faster than before.

The caveat is the story

Zettair is also an old name. An archived RMIT University project page described the Zettair search engine as a compact text search system from RMIT's Search Engine Group for indexing and searching HTML or TREC collections. A TREC paper from 2004 by RMIT authors including Hugh E. Williams documented Zettair in evaluation settings.

That provenance cuts against the easy social-media version of the story. Williams did not discover information retrieval by chatting with Claude Code. He brought the old system, the evaluation instincts and the product taste. Claude Code supplied the execution layer around those decisions.

The distinction is important for founders deciding how much to trust agentic coding tools. If the claim is "AI can build a search engine," the right response is skepticism. If the claim is "AI can help an experienced search engineer rebuild and modernize a known search system into a working product," the response is operational: what parts of the stack can now be delegated, what review loops are mandatory, and how much faster does the expert move?

Anthropic's product bet is that builders stay in charge

Anthropic describes Claude Code as an agent that understands a codebase, edits files, runs commands, and integrates with existing development tools. It is positioned for routine tasks like bug fixes, tests, refactors, and feature work.

Business Insider's August 2025 write-up of Williams' earlier Claude Code experiment is a useful predecessor. In that case, Williams used Claude Code to build an AWS-based system in 48 hours, a task he said would normally take at least three weeks. The same piece reported hard limits: conversation context compression broke things, a cleanup request wiped out important work, and Williams learned to back up code and reset conversations around milestones.

The throughline between the August 2025 AWS experiment and the June 2026 Zettair project is not that agents became autonomous engineers. It is that Williams changed the shape of the work. He was still setting direction, inspecting output and supplying the technical frame. Business Insider says Williams' latest takeaway is that building with AI feels less like programming and more like coaching, and that experienced engineers still make the best coaches.

What founders should take from Zettair

The Zettair experiment lands at an awkward moment for software teams. Investors and executives want to believe coding agents will compress headcount needs. Engineers know that shipping code has never been the whole job. Architecture, state management, testing, observability, latency, cost control, security and product judgment sit outside the raw act of generating files.

Williams' project supports the more durable middle view. Coding agents can reduce the distance between idea and working system when the operator knows the terrain. They can revive older systems, wrap them in modern interfaces and automate the work that would previously consume days of scaffolding. But the agent still needs a coach who can recognize when a search result is bad, when a snippet is misleading, when an index design will not scale, or when an AI-generated summary creates more liability than utility.

The open questions around Zettair are the same questions a buyer, investor or technical founder would ask before treating the demo as proof of production readiness. Business Insider reports the index size and feature set, but not latency, infrastructure cost, relevance benchmarks, human evaluation results, uptime, repository history or deployment details. There is also no evidence in the available reporting that Zettair is a company, a commercial product or a funded startup.

That does not make the experiment small. It makes it precise. A search veteran took a system with roots in early-2000s information retrieval and used a modern coding agent to turn it into something that looks closer to a contemporary search product. The result is not a tombstone for engineers. It is a map of where experienced builders are likely to get leverage first: in domains where they already know enough to tell the machine what good looks like.