← All speakers

Thariq Shihipar

Conference affiliation: Claude Code · Anthropic · 2026

Thariq Shihipar is an independent engineer who worked on Claude Code at Anthropic. He left Anthropic in October 2026 after more than a year on the team. His contributions included AskUserQuestion, the tool that presents structured questions when an agent needs to clarify a user's requirements. Before Anthropic, Thariq co-founded scholarly publishing platform PubPub and led gaming company One More Multiverse. His current work explores how tools and interfaces let agents accomplish tasks—and help people understand the decisions made along the way.

On this page

From notifications to collaborative publishing

Thariq began building companies as an undergraduate at the University of Toronto. With Fahd Ananta and Guru Mahendran, he co-founded Chime, a Chrome extension that gathered notifications from email and social services into one stream. HubSpot acquired the startup in 2013. Thariq subsequently worked there as a software engineer, helping establish its Signals sales-software team.

He entered graduate study at the MIT Media Lab in 2014 and continued as a research scientist after his master's work. His projects included NewsClouds, a news-visualization project, and PubPub, which he co-founded with Travis Rich. Their collaborative publishing platform let research evolve on the web: readers could discuss and annotate work, authors could revise it, and publications could incorporate interactive material. The team worked with MIT Press on journals and books, connecting that approach to publishing workflows.

Thariq later developed products for clients at Rocket Insights and built Edgeout.gg, a gaming analytics platform for professional League of Legends teams acquired by Blitz. His career history describes building a memory-injection tool to extract data from active games for analysis.

Building worlds and exploring models

In 2019, Thariq co-founded One More Multiverse with Hisham Bedri and illustrator Sara Alfageeh. The company joined Y Combinator's Winter 2020 batch and developed a browser-based game-creation platform inspired by tabletop role-playing games and Roblox. Illustrated worlds and creator-made assets gave players tools to build and share games. As founder and CEO, Thariq worked across engineering, product, design, art and marketing, managing a team of more than 30. He describes the platform as reaching more than a million users.

His subsequent AI experiments included agent creation, copyediting, email prioritization and assisted world building. During a Goodfire research fellowship from December 2024 to March 2025, he worked on automated interpretation of model features and tools for visualizing them. He joined Anthropic in June 2025, while his listed Multiverse tenure continued through October that year. His account of the transition describes winding down One More Multiverse and finding Anthropic more like a large startup than he had expected, with many problems he could take on.

Two technical essays explain questions he was exploring before joining Anthropic:

  • Model steering and its limits. In Should Developers Care about Interpretability?, Thariq considers whether manipulating internal model features could give developers finer control over style, preferences and classification. He also examines the risks: misunderstood features, excessive steering and interactions between features can produce unwanted behavior. He presents these applications as possibilities to investigate rather than established solutions.
  • Semantic sorting with TrueSkill. His LLM-powered sorting method asks a model to compare small batches, then uses TrueSkill to accumulate ratings and uncertainty estimates. This separates semantic judgment from ranking bookkeeping and allows new items to enter without repeating every comparison.

An environment where agents can investigate

Thariq favors giving agents familiar computing tools and inspecting how they use them. In his Claude Agent SDK workshop, he explains why Bash and the filesystem matter beyond coding: an agent can compose APIs, generate scripts and preserve intermediate results instead of doing every operation inside its conversational context.

An email-spending question makes the mechanism concrete. Rather than asking a model to mentally total a pile of ride-sharing receipts, an agent can save search results, extract amounts, calculate a total and revisit the original messages to check what each amount represents. His Pokémon prototype uses the same approach to search competitive-game data and generate scripts that analyze teammates and moves.

That flexibility has costs. Bash requires the agent to discover commands, which adds latency; generated code may need compilation and checking. Thariq still favors dedicated tools for actions such as sending an email or editing a file when the application needs to recognize the action and present it for approval. His argument is for choosing the interface around the work, rather than making every operation a separate tool.

Progressive context disclosure extends this approach to organizational knowledge. His writing on skills treats a skill as a discoverable folder of instructions, references, scripts and resources. Its most useful content often captures recurring mistakes: which database field actually identifies a user, which record version to select, or why a successful HTTP response does not prove that a payment was processed. Reusable helper code lets an agent compose existing functionality, while references can be loaded when relevant. He advises supplying goals and constraints without prescribing steps that prevent adaptation.

Thariq starts verification with checks that produce actionable feedback: compilation, linting, assertions and rules that reject an edit if the agent has not first read the file. He distinguishes these deterministic checks from review by another model, which he recommends starting with fresh context to avoid carrying over the original agent's assumptions. Sandboxing, scoped credentials and command inspection add protection. Reversibility also affects which tasks he considers suitable for autonomy: restoring a Git revision is simpler than recovering from an action that has changed an external system.

Asking better questions as models improve

Thariq worked on AskUserQuestion when he first joined Claude Code. In his Field Guide to Fable, he describes how its usefulness changed across model generations. Opus 4 needed considerable adjustment to call the tool; with Opus 4.5, he could ask the agent to interview him about a specification. Later models could build HTML reports with questions embedded alongside the information needed to answer them. The same interaction became more useful as the model became better at eliciting requirements.

His techniques address different gaps in a specification. A blind-spot review can investigate an unfamiliar authentication module before implementation. Contrasting prototypes can reveal visual preferences that a user would struggle to put into words. Implementation notes record decisions the agent made when it encountered something unspecified. Thariq also asks agents to quiz him about their changes so he can explain the work when submitting or merging it.

He revisits prompting assumptions too. Examples that helped earlier models can constrain more capable ones; he describes the Claude Code team moving toward fewer instructions and more context for its newer models. His public illustrated explanation skill carries the interest in understanding beyond programming, using HTML artifacts with large pictures and few words to introduce unfamiliar topics.

A founder's changed sense of possibility

Thariq connects his enthusiasm for coding agents to the constraints he experienced running a company. Features and performance improvements once competed for weeks of engineering time. Revisiting his former company's codebase with agents, he found some of that work possible in hours. He argues for testing whether familiar trade-offs still hold rather than accepting them before attempting the work.

That optimism includes a sense of loss. He describes loving handwritten code and the experience of holding a codebase in his mind, alongside prolonged debugging and failed projects. His aim is to attempt more ambitious work and spend more time with people he cares about. He also preserves a founder's distinction: making software easier to build does not make it easy to discover what creates value.

Talks by Thariq Shihipar

3 talks

Loading more talks and highlights…

References