Thariq Shihipar is an engineer working on Claude Code at Anthropic. His earlier projects span game development, academic publishing and small AI tools, including One More Multiverse and PubPub.[1][2]
Start with LLM-Powered Sorting with TrueSkill →
X · trq212 Background
MIT lists Shihipar among its 2016–2017 master’s graduates in Media Arts and Sciences. While at the Media Lab he developed PubPub with Travis Rich, providing an open authoring and publishing platform. His personal portfolio also lists Chime, a startup acquired by HubSpot, and Edgeout.gg, a gaming-analytics product sold to Blitz. These are separate projects from his later work on coding agents.[1][2][3]
Career
Shihipar says he ran the Y Combinator-backed gaming company One More Multiverse for five years. His portfolio describes it reaching more than a million users, a figure reported by its founder. He now works on Claude Code at Anthropic. His small AI projects include Quick Edit for copyediting and prototypes for email prioritization and role-playing world creation, giving his writing concrete product problems to examine.[1]
Contributions
His sorting essay addresses a practical limitation: asking a language model to order a large list can exceed its context or output limits and produce inconsistent rankings. He proposes having the model compare small batches, then updating a TrueSkill rating for each item. Repeated comparisons produce a global ordering with uncertainty estimates, and new items can enter without re-sorting the entire set at once. This adapts an existing rating method to semantic comparisons; it is a proposed implementation technique, not evidence that arbitrary rankings become objectively correct.[4]
Ideas
In his interpretability essay, Shihipar asks how inspecting and activating internal model features might help developers control style and behavior. He separates the promise of steering from existing prompting and training techniques and discusses possible applications. It is a developer-oriented argument about research possibilities, rather than a claim that he invented the underlying interpretability methods.[5]
Tools, writing & talks
Writing, research & talks
Start here
LLM-Powered Sorting with TrueSkillthariq.ioA method using small model comparisons and probabilistic ratings to rank larger collections.
Begin with the sorting walkthrough for a concrete engineering problem and implementation sketch. Python familiarity and a basic understanding of model prompts help; TrueSkill is introduced in the article.
Author[4]An Anthropic engineer says Claude Max plans now include monthly Claude API credits matching the plan price, $100 or $200. They demo a personal new-tab homepage generated daily from sites they read. The demo video was not readable.
Catalog workOriginal sourceAnthropic released Claude Haiku 5.5, its small model, with a sharp price cut: about 10x cheaper than Haiku 4.5 under 100k tokens and roughly 75% cheaper on average. Available for computer use, workflows and the API.
Catalog workOriginal sourceLatent Space talks with Anthropic's Thariq Shihipar about Claude Code's next phase, alongside a recap of Anthropic's recent releases including Sonnet 5.5, Opus 5.5, Claude Mods, and the plugins portal.
Catalog workOriginal source- The Future of Claude Code: Mods, Mutable Software, & Multiplayer Agents — Thariq Shihipar, Anthropicyoutube.com
Anthropic's Thariq Shihipar covers Claude Code power-user workflows, Mods for rewriting the harness, multiplayer agents, mutable software, and agent security incidents with defenses like sandboxing, classifiers, and probes.
Catalog workOriginal source Anthropic's Thariq breaks down what 'effort' actually controls in Claude models: more verification and edge-case testing at high effort, faster iteration at low/medium, illustrated with Terminal-Bench 3.0 results and a practical build loop.
Catalog workOriginal sourceClaude Code adds `claude plugin eval init`, so developers can define test cases, run a plugin or skill against them with and without the plugin, and score the delta to catch regressions when new models ship.
Catalog workOriginal sourceA slash command that turns an unfamiliar module or incident into a picture heavy explainer.
Catalog workOriginal sourceAnthropic is launching enterprise safeguards for Fable class models that run on a customer's own infrastructure, giving teams control over data location and access, with wider rollout planned for the fall.
Catalog workOriginal sourceClaude's text watermark changes where sampling randomness comes from rather than which words get picked, which is why output quality is unaffected. Anthropic is rolling it out for EU AI Act compliance; an interactive explainer walks the mechanism step by step.
Catalog workOriginal sourceAnthropic is embedding invisible watermarks across everywhere Claude is offered. The mark lives in the text rather than in metadata, travels with copy-paste and can persist through some editing, with a detection API planned — enough to ask whether a PR came from Claude Code.
Catalog workOriginal sourceAnthropic's Thariq frames Claude's text watermarking as EU AI Act compliance work that other labs are also adopting, with a public text-detection API coming.
Catalog workOriginal sourceAnthropic's Thariq argues auto mode is safer than manual review of individual commands, announcing its default rollout with no added latency cost for the classifier.
Catalog workOriginal sourceA user observation that connecting a Claude Connector (Gmail, Calendar, Slack, etc.) to Claude also makes that connection available within Claude Code, including inside Artifacts, a behavior many users may not realize.
Catalog workOriginal sourceA write-up from the Claude Code team on stripping roughly 80% of the system prompt for newer models, and what that exercise taught them about writing system prompts, skills, and CLAUDE.md files for more capable coding models.
Catalog workOriginal sourceAn article by Thariq exploring how to work effectively with Claude's Fable system, drawing on the insight that 'the map is not the territory.' It guides readers on understanding the gap between their prompts/context and the actual work being done by AI. Useful for developers and creators trying to get more out of AI-assisted workflows.
Catalog workOriginal sourceThariq makes the case for switching from Markdown to HTML as the default agent output format. Argues HTML is richer (tables, SVG, CSS, interactivity), easier to share, and keeps you more in-the-loop with what Claude is generating — even though it's 2-4x slower to produce than Markdown.
Catalog workOriginal sourceAn update to Claude Code's /usage view, shaped by how differently people manage sessions, and what the move to a one-million-token context changes about when to start a new one.
Catalog workOriginal sourceA developer-oriented explanation of model features, steering and possible applications.
Author[5]