Skip to content
View Oldcircle's full-sized avatar
  • Central China Normal University
  • 中国

Block or report Oldcircle

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Oldcircle/README.md
Yuan Bo (@Oldcircle): I build AI agents that show their work.

Homepage 中文主页 AACL 2026 paper

I make agent skills, developer tools and browser games. The ones I like best refuse to guess: a claim has to name the file it came from, observations are sealed before any reasoning starts, and every number is run before it goes on the page. AI master's student at Harbin Institute of Technology, Shenzhen.

Selected work

geo-sleuth: photo geolocation skill for AI agents. 27,335 → 171 → 22 → 1, ±2 m

🧭 geo-sleuth  GitHub stars

Finds where a photo was taken, and shows its work. No text, no plates, no landmarks: one bridge and one mountain put the camera within 2 m.

npx skills add Oldcircle/geo-sleuth

Holmes: show it a photo, it tells you things you never said

🔍 Holmes 福尔摩斯

Reads a photo the way Sherlock Holmes reads a stranger. Observations are sealed before any deduction; on a real desk photo the owner graded 5 hits, 1 miss, 0 made-up observations.

npx skills add Oldcircle/holmes-skill

Both are plain Agent Skills and work in Claude Code, Codex, Cursor, Gemini CLI, OpenCode and GitHub Copilot.

Tools for agents

Project What it does
🪴 Trellis The discipline layer for file-based agent memory: budgets, ownership rules, and a linter that keeps CLAUDE.md and MEMORY.md from rotting.
🎓 deep‑lecture A teaching method as a Claude Code skill: read the primary sources, run every number, publish one interactive page, patch it wherever the learner gets lost. In Chinese.
🪟 wsl‑agent‑kit One double-click on Windows sets up WSL2, your pick of nine AI agents and a Chinese office workspace. In Chinese.
⚡ ai‑shell Build your own Claude Code from scratch: streaming query loop, tools, permissions, context compaction.
🔬 trace‑viewer Every LLM call, tool call and token spent in an OpenClaw agent run, laid out step by step.

Research

RouterKGQA: a specialized LLM drafts the reasoning path, a general LLM repairs it when the path is unreachable

RouterKGQA: Specialized–General Model Routing for Constraint-Aware Knowledge Graph Question Answering
Bo Yuan*, Hexuan Deng*, Xuebo Liu, Min Zhang · AACL-IJCNLP 2026, Main Conference · *equal contribution

A small specialized model drafts the reasoning path over the knowledge graph. A large general model is called in only when that path cannot be reached: +3.57 F1 over the previous best, at 1.15 LLM calls per question.

arXiv:2603.20017 · PDF

Simulations and visual explainers

backprop-viz: the loss chapter
backprop-viz
A 3D walkthrough of a GPT's backward pass, built on llm-viz. Every block holds its gradient, computed live.
Anima: a pixel-art seaside town
Anima
Seven anime characters, each its own LLM agent, in one seaside town. Friendships and grudges come out of the interactions.
BlockFire: a voxel forest and a zombie
BlockFire 方块火线 · play ↗
Minecraft meets first-person shooter, in one HTML file. Dig, build walls, hold off the waves.

Also: Stardew Anima, an MCP server with 28 intent-level tools that let an LLM play Stardew Valley without screenshots.

Games you can play in the browser

Xiangqi
Xiangqi · Chinese chess vs an AI
Zen Gomoku
Zen Gomoku · five in a row vs an AI
Joker Poker
Joker Poker · a Balatro-style roguelike
Survivor Royale
Survivor Royale · survive, hunt, evolve
Neon Survivor
Neon Survivor · twin-stick zombie shooter
Alpine Escape
Alpine Escape · an endless ski run
Holiday Memories
Holiday Memories · photos as a 3D tree
Lumina
Lumina · put the gradient back in order

Upstream

Earlier projects
  • StoryForge: one sentence in, one comic or short-drama episode out
  • ChromaCheck AI: colour palettes from a description, with Gemini, DeepSeek or OpenAI
  • paper-tl: upload a paper PDF, read a Chinese translation laid out for a phone
  • LinkStream: a phone and PC sync assistant

oldcircle.github.io · 中文主页

Popular repositories Loading

  1. geo-sleuth geo-sleuth Public

    An agent skill that finds where a photo was taken — OpenStreetMap geometry, elevation skylines, satellite imagery and street view — and shows its work. Works with Claude Code, Codex, Cursor, Gemini…

    Python 1.6k 156

  2. holmes-skill holmes-skill Public

    An agent skill that reads a photo the way Sherlock Holmes reads a stranger (福尔摩斯): observe first, deduce second, every conclusion cites its evidence and you grade it. Works with Claude Code, Codex,…

    TypeScript 11 1

  3. RouterKGQA RouterKGQA Public

    Python 4 3

  4. Xiangqi Xiangqi Public

    中国象棋AI版本

    TypeScript 2 1

  5. Oldcircle Oldcircle Public

    Profile README

    2

  6. Alpine-Escape Alpine-Escape Public

    滑雪大冒险

    TypeScript 1