Writing
AI, engineering, and method. Writing down what I've figured out — and what I haven't.
2026
16- 03-18Optimizing AI Workflows with "Pruning" Thinking: A TikTok Video Summarization ExampleIn computer science, this is called pruning. Through a complete iterative process of building a TikTok video summarization tool, learn how to optimize AI workflows to the extreme using pruning thinking.4 min read
- 03-01Nine Principles After Reviewing the Hottest Prompt Engineering PracticesNine efficient principles summarized from four prompt engineering whitepapers: from efficiency formulas to pipeline splitting, from XML tags to parameter configuration—transforming prompts from one-time conversations into reusable assets.8 min read
- 02-26Nano Banana Version Logic: Gemini Image Series Naming DecodedExplaining the version numbering logic of Google's Gemini Image series (Nano Banana), and why Nano Banana 2 didn't surpass Nano Banana Pro.4 min read
- 02-08Seven Tips for Using Claude CodeFrom Plan mode to Sub-fractal Architecture, from MCP retrieval to Agent collaboration—seven battle-tested tips to double your Claude Code efficiency.8 min read
- 01-30Five Insights from Altman: AI, Entrepreneurship, and the FutureAfter watching OpenAI CEO Altman's live interview: the future of software engineers, GTM bottlenecks, competitive moats for AI products, GPT-5 writing issues, and the most critical skills in the AI era.5 min read
- 01-26Why Skill Quality Differs So Dramatically Between PeopleWriting a Skill requires planning first. Using daily digest scraping as a case study, this article breaks down how technical research can create order-of-magnitude improvements in Skill stability, speed, and Token consumption.4 min read
- 01-22Troubleshot All Day, Finally Restored Proxy Speed to Gigabit LevelFrom selecting a Clash distribution to tuning Tun stack mode, I troubleshot the proxy speed issue step by step and finally pinpointed the performance bottleneck caused by DNS resolution going through gVisor.4 min read
- 01-20X's Official Open-Source Recommendation AlgorithmX officially open-sourced the x-algorithm project, completely exposing the core code of its recommendation system. An in-depth analysis of positive/negative weight parameters and Grok transformer-driven recommendation mechanism.4 min read
- 01-16Roll Your Own Budget Typeless — 1/10 the Cost, 85% of the EffectVoice recognition post-processing prompt V2 — implemented with GLM-4.7 for intelligent ASR cleaning, aggressively strips filler words and fixes mixed Chinese-English terminology.5 min read
- 01-14Stop Memorizing Skill Docs — Start BuildingA practical guide to learning Skill: from setting goals to understanding concepts to learning by doing. Step-by-step guide to creating automation Skills that actually improve your productivity.6 min read
- 01-13Dan's Take on the Most Important Skill for the Next Decade: AgencyAn exploration of Dan's article on why 'agency' is the most critical skill in the AI era, and how to reflect, iterate, and act like an Agent.6 min read
- 01-12Skill = Agent + Workflow: The Underlying Logic from Anthropic's Five Patterns to SkillReviewing Anthropic's "Building effective agents" and understanding why Skill is the fusion of Agent and Workflow.5 min read
- 01-10The Boundary Between MCP and SkillMCP provides remote service capabilities; Skill is a local Agent SOP orchestrator. Understanding their boundaries helps you choose the right tool for the right job.4 min read
- 01-07What Every Powerful AI Agent Should Include: A Deep Dive into Four Design PatternsFrom delegation to state machines, this article breaks down four common design patterns in AI Agent design: Orchestrator-Workers, Pipeline, Circuit Breaker, and State Machine.6 min read
- 01-06The Four Elements of AI Agents: From Large Language Models to Intelligent AgentsLLM + Memory + Planner + Tool Use = AI Agent. A breakdown of Agent components, the ReAct design pattern, and why the same model can produce vastly different results across different products.7 min read
- 01-03An Absolutely Underrated Open-Source Gem: OpenCode + oh-my-opencodeA free, open-source AI coding combo that combines all the strengths of Claude Code and AmpCode, even surpassing them in certain aspects.5 min read
2025
4- 12-27PACER Learning Method Explained: A Cognitive Upgrade from "Hoarding" to "Digesting" KnowledgeAn breakdown of Justin Sung's PACER system, explaining five types of knowledge from a neuroscience perspective and their corresponding efficient learning strategies.5 min read
- 12-22I Never Come Up with Prompts for Image GenerationSharing a complete workflow for generating images with Gemini: let AI do the planning first, generate specialized prompts, then iterate and optimize the results.4 min read
- 12-15Dissecting the 'Life K-Line' Underlying Logic: BaZi Fortune-Telling and Decade CyclesWith background in both trading and metaphysics, I provide a detailed analysis of the Four Pillars BaZi theory behind the "Life K-Line" website, its decade cycle system, and the crypto industry propaganda mixed into the interpretations.5 min read
- 12-02Why Enabling TUN Mode Bypasses Geo-Restrictions: A Deep Dive into Proxy MechanicsFrom OS-level network proxies to TUN virtual interfaces, this article explains the fundamental differences between proxy mode and TUN mode, and why TUN mode is better at concealing proxy status.4 min read