Blog
Thoughts, tutorials, and technical writings.
OpenClaw PR #86261: Plugin Skills Sandbox Sync Fix
Plugin skills couldn't be read by agents in sandbox mode because symlink syncSourceDir was unset. Fix earned diamond lobster rating.
Claude Code TEAM Mode and Migration Guide
TEAM mode enables multi-agent collaboration with effortLevel inheritance. When migrating, backup plugins/cache/ directory or plugins will show enabled but not work.
Game Booster Safety Toolkit: Audit, Block, One-Click Toggle
Technical audit of a game booster revealing kernel drivers, full NIC binding, self-signed CA certificate, and 22 ad SDKs. Developed a toggle tool that blocks all system capabilities when not in use.
OpenClaw PR #85850: A Superseded Successful Contribution
A merged PR that later had better alternatives. Learning notes on multi-layer defense, avoiding broad rules, and understanding complete data flow.
OpenClaw PR Aftermath: Second Closure and Microsoft VP's Public Apology
Submitted refactoring PR #82683, closed again. Timestamp evidence led to Microsoft VP's public apology. Full retrospective.
OpenClaw PR Closed: Passed All Reviews But Overwritten
Independent contributor passed all 8 rounds of ClawSweeper review, but PR was closed by Bot in favor of a later maintainer PR with unresolved defects. Reflection on open source governance transparency and contributor rights.
Recruit Skill - Agent Persona Builder
From 214 lines down to 16: a Claude Code Skill that builds agent personas through iterative questioning.
Vibecoding Interview - Satirical Interview Roleplay
A fun interview roleplay game where AI plays an HR who's not quite up to speed on AI coding tools.
Compose Web: VS Code Plugin-Driven Full-Stack Scaffold System
Deep dive into VS Code plugin-based full-stack website scaffold design—three-layer block structure, natural language driven, multi-stack support with self-verification and self-healing mechanisms.
ECHO_CLAW: Context-Driven Game Agent Framework
Deep dive into ECHO_CLAW's design philosophy—achieving intelligent behavior through context + skill invocation, similar to Claude Code CLI's architecture, extensible to any game.
ECHO_loop: Engineering React Agent Loop for Minecraft AI
Deep dive into engineering React Agent Loop for Minecraft AI companion, including Electron desktop architecture, LangChain Agent orchestration, Mineflayer game interaction, and local memory system design.
Mirror: The Path to True Intelligence
Exploring the implementation of 'true intelligent entities'—neural-symbolic hybrid architecture, unified memory, dream engine, independent awakening process, pursuing non-tool-like intelligent existence.
SuperHarness: AI Agent Development Framework Benchmarking Claude Code
SuperHarness is a production-grade AI Agent framework built with Rust + Python, benchmarking Claude Code and LangChain. Phase 0 complete with 192 tests passed.
OpenClaw PR #80773: config unset Dry-run Feature Implementation
Adding --dry-run, --json, and --allow-exec flags to openclaw config unset command, passing ClawSweeper eight-round review. Includes SecretRef fallout detection, detailed use cases, and review process analysis.
Terminal Graphics Residue Investigation and Issue Submission
While using Claude Code for long conversations, discovered abnormal terminal display after context compression—old content remained as static bitmaps, overlaying current output.
Complete Guide to Cleaning Sangfor EasyConnect Residue
Documenting a complete process of discovering and thoroughly cleaning Sangfor EasyConnect VPN client residue, including background processes, system services, root certificates, NSP drivers, and other stubborn remnants.
Claude Code Ghost Session Bug: Unknown Source stdin Injection
Found conversation records in resume list that were never input, discovered abnormal queue-operation behavior through JSONL log analysis.
Disable Claude Code's Auto Prompt Suggestions
Claude Code by default shows 'Try...' suggestions after conversations. If you find these distracting, you can disable them through configuration.
Claude Code Windows Symlink Permission Issue Investigation
From background sub-agent showing 0 tokens to discovering Windows symlink permission issue—a complete bug investigation process.
Switch to MD: Making AI Agents Understand All File Formats
Convert any file format to Markdown, enabling AI Agents to perfectly understand all content. One line of code, supporting cloud & local models.
Automation Scaffold: Unified Monorepo Automation Collection
A unified Monorepo automation project collection including AI tables, RPA, data scraping, and Agent subsystems.
WeChat Mini Program Cloud Showroom: AI-Powered Product Display Platform
A WeChat mini program DEMO for clothing product cloud showroom with product display, VIP permission control, and AI image generation.
Minecraft Lohse Skin: Procedural Pixel Art
Procedurally generating Minecraft skins using Python PIL, creating pixel art based on the character Lohse from Divinity: Original Sin 2.
ECHO_feeling: Emergent Minecraft AI Companion
A desktop application for emergent AI companion in Minecraft, based on cognitive science architecture with independent in-game operation and long-term memory.
Building an AI Project Context Skill
A project context injection tool for AI-assisted coding scenarios, designed to reduce repetitive explanations in development workflows.
Mind Memory Lab: Memory-Oriented Agent Framework
Memory-oriented agent framework prototype validating long-term memory and dialogue continuity, with pluggable LLM adapters and SQLite FTS full-text indexing.
HR Interview System: AI Resume Analysis and Interview Platform
An AI agent-based HR interview and candidate analysis platform supporting intelligent resume scoring, AI interview dialogue, and in-depth analysis.
OKR Coach System: Intelligent Scoring and Training Platform
An AI agent-based OKR scoring and training system supporting file upload, automatic cleaning, intelligent scoring, and result export.
Neural Attention Pattern Visualization Implementation
A hands-on guide to understanding and visualizing attention mechanisms in transformer models with code examples.
Kenshi Wiki Data Scraping: Anonymous Crawling and Data Cleaning
Systematic scraping, cleaning, and organizing of Kenshi game Wiki data to build structured game knowledge datasets.
Reverse Engineering Kenshi: DLL Injection for Multiplayer
How I implemented multiplayer functionality in a single-player game through DLL injection and memory hooking, without modifying game binaries.
Kenshi RNG Fix: Solving Fixed Random Seeds via DLL Injection
A memory fix tool for Kenshi's fixed random number seed problem, replacing the game's runtime random seed through DLL injection to achieve randomization differentiation for each new game.
Running Redis Clusters on Kubernetes: Technical Notes and Architecture Survey
A study of approaches for deploying stateful Redis on Kubernetes.
Building a Multi-Stage Decision System with LangGraph
How I designed a geological hazard assessment system using LangGraph's multi-stage decision architecture for intelligent risk evaluation.
Consciousness Kernel: A Computational Model in Racket
A computational model approaching consciousness based on Global Workspace Theory, using Racket's metaprogramming capabilities for exploratory implementation.