Skip to main content
Zhenyu He · Jobs Stroustrup
English 中文

Department Head @ ONE-AGI

Team project

Basic Info

  • Period: 2025 — Present
  • Role: Department Head (AGI Division Lead)
  • Organization: ONE-AGI; the AGI Division of UC Berkeley BETA (Berkeley Emerging Technology Association)
  • Location: UC Berkeley

Role Summary

ONE-AGI is Zhenyu’s “department-of-one for AI-powered personal effectiveness,” initiated at UC Berkeley in 2025. The framing is literal: treat the practice of using AGI for individual productivity as a department with its own discipline, artifacts, and review cadence; engineer every recurring task into a workflow an agent can complete safely; turn “10x productivity through LLMs” from a slogan into shipped artifacts. The same role doubles as convener of UC Berkeley BETA’s AGI Division, running weekly cross-disciplinary seminars on AGI safety, AI policy, and human-AGI coexistence.

Concrete Artifacts

  1. A Claude Code Skill authoring methodology. After Anthropic released Claude Code Skill in late 2025, Zhenyu independently developed a paradigm built on five tenets: phased explicit workflows, data-and-logic separation, human-in-the-loop on high-risk actions, an examples-write-back data flywheel, and entity-organized long-term memory. Captured as a reusable methodology page that grows with each new skill.
  2. private personal automation. The first full instance of that methodology: a Claude Skill that automates browsing, liking, and commenting on WeChat Moments. Combines AppleScript (bring the desktop app forward), the computer-use MCP (screenshot and click), and Claude’s vision plus Chinese generation, layered over five data files (persona, examples, blocklist, like-only, friends). Every comment is presented as a draft for user confirmation; the confirmed final form is appended to the examples corpus, producing a personalization flywheel without any model fine-tuning.
  3. MyMemory wiki. A Karpathy-style LLM Wiki launched in April 2026 as the canonical store for the writer’s personal record: sources, entities, concepts, experience, skills, meta. The LLM is the maintainer; the human reviews and approves; per-page publish flags govern what ever leaves the local store.
  4. A three-layer fail-closed publishing pipeline. The very website you are reading is itself an ONE-AGI artifact: a MyMemory canonical layer, an auto-generated MyMemoryForAstro middle layer with hard exclusions and section-level redaction, and an Astro-rendered public layer with schema-level fallback. Architecture distilled from nineteen Opus subagent research briefs and a Wave-based decision protocol.
  5. AGI-issue seminars. Weekly Berkeley discussions on AGI safety, policy, legal frameworks, and human-AGI coexistence; occasional public workshops on practical AI tooling and prompt design for newcomers.

Connection to the Broader Research and Builder Identity

ONE-AGI is not a separate track from the climate-physics research at Romps Lab or the embodied-data company (now passive shareholder). They are the same identity expressed on three surfaces: research treats large physical systems with LES and inverse methods, the startup folds those methods into a data flywheel, and ONE-AGI folds the same engineering discipline back onto the practitioner’s own productivity surface.

Timeline

Began 2025; ongoing as of 2026-05; both the BETA seminar cadence and the personal skill library remain active.

Skills Used