Blog
Modeling a Game Character with Codex + Blender: A Step-by-Step Walkthrough
Using the Blender MCP server to let Codex drive Blender directly — installing uv, wiring up MCP, breaking a character down into a verifiable feature checklist, and accepting the final asset by mesh count, structure, materials, scale, and file format.
I Built a 3D Game You Can Build, Shelter, and Defend In with GPT-Astra + Three.js
From a one-line idea — a woodland building game — to a fully playable browser world: how I used GPT-Astra as a continuous design, implementation, testing, and fixing partner to build Wildwood, complete with construction, furniture, day/night weather, pathfinding, monster raids, and local saves.
From Vibe Coding to Spec Coding: Building a Project-Level AI Workbench with Trellis
Vibe coding is great for throwaway scripts, but left unchecked it turns real projects into unmaintainable messes. Trellis practices spec coding by persisting project rules, task progress, and working memory into .trellis/, fixing session amnesia, tool lock-in, and the lack of a project-level feedback loop.
【DeepSeek Harness Plugin】01: Writing Your First Plugin from Scratch
Write a DeepSeek Harness plugin from scratch: teach the assistant to greet someone with 'Hello, XXX!'. Through this minimal example,拆解插件是什么、怎么写、怎么跑起来。
From Vibe Coding to Vibe Worlding: AI Is Learning to Build Its Own 3D Worlds
HKUST and Tencent present VibeWorlding, a multimodal RL framework that lets AI agents autonomously construct and edit interactive 3D worlds through natural language — with open-source models already surpassing GPT-5.5.
Vibe Coding Best Practices: From 'Let AI Write Code' to a Verifiable Software Engineering Loop
Vibe Coding in 2026 is no longer just prompt tricks — it's a four-layer engineering system: Prompt, Context, Harness, and Loop. From 'tell the agent what to do' to 'tell it what counts as done,' and on to TDD, subagents, skills, harnesses, and review loops — a survey of the practices that turn AI-written code into a sustainable software engineering system.
Dissecting Grok Bot: Long-Term Agents, Async Collaboration, and Layered Memory
Grok Bot's agents carry identity, long-term state, memory, tools, and a run queue. Multiple bots coordinate via async messages or sequential Group discussions — no Supervisor model needed. Its context is layered, not infinitely appended.
DeepSeek Harness (Evaluation): Can It Replace Claude Code? Don't Just Look at the Model
Series #5: An honest evaluation of whether DeepSeek Harness can replace Claude Code — covering what each actually does, where they overlap, and when to choose which.
GLM-5.3 Flash (Ox Alpha) Launches: Opus 4.8 Performance at 1/40th the Price
Anonymous model Ox Alpha tops OpenRouter leaderboards — then its identity is revealed: GLM-5.3 Flash. Priced at 1/40th Opus 4.8, with matching capability. Three frontend coding benchmarks put it to the test.
UI Skills: 181 Design Skills for AI That Actually Look Professional
UI Skills is a design-engineering AI agent skill library with 181 specialized skills — from anti-gradient constraints to full accessibility audits — compatible with Claude Code, Cursor, Codex, and more.
DeepSeek Harness (Architecture): Why Even the Agent Loop Is Replaceable
Most plugin systems keep an immutable core loop. DeepSeek Harness turns even the Agent Loop—model decision, tool execution, result feedback—into a replaceable plugin. That's the real difference behind 'Everything is a Plugin'.
DeepSeek Harness Quickstart: From Installation to Your First Request
Get DeepSeek Harness running in four steps: install, configure a model provider, add a workspace, and send your first message.
DeepSeek Harness, Explained: It Looks Like Claude Code — So Why Call It a Harness?
A naming-first tour of DeepSeek Harness: why it feels like just another coding agent, yet the team insists on calling it a Harness. Splitting model, execution layer, and agent product into three layers to see what "Everything is a Plugin" actually changes.
Engineering Vibe Coding: A Five-Layer System for Production-Grade AI-Generated Code
How the author turned raw AI coding into a five-layer Claude Code system of rules, process, memory, skills, and collaboration, cutting review time 40 to 8 minutes.
Claude Code Sessions Can Now Talk to Each Other: A Hands-On Guide to v2.1.224's Cross-Session Messaging
Claude Code v2.1.224 ships ListAgents and SendMessage for real-time session coordination. A hands-on WSL walkthrough: two sessions, a live demo, cross-machine limits.
An Agent's Ceiling Isn't the Model — It's Your Team's Knowledge: A Working Knowledge-Flywheel Playbook
When agents drive frontline work, knowledge supply breaks: bad retrieval, stale content. A playbook for building a team knowledge flywheel — goals, build steps, anti-patterns.
I Ported Hermes's Core Engine into Obsidian — The Results Are Stunning
I ported Hermes's core engine — memory retrieval, skill marketplace — into Obsidian. The result, obsidian-cc, gives your 2000 notes a real digest-deposit-grow loop.
94% AI-Generated Code: How We Ran the Full Feature-Development Pipeline Through a Single Skill
8 semantic stages, a red-line rule system, a three-tier knowledge base, and cross-session TECH_SPEC.md let AI run enterprise features end-to-end at 98% code-generation.
DeepSeek Harness Developer Preview: Everything Is a Plugin
DeepSeek Harness Developer Preview (v0.1) is open under MIT: models, tools, skills, sessions, sandboxes, loops, UI — all plugins. The four run modes and how to start.
OpenClaw: Core Architecture, How It Works, and Agent Deployment
OpenClaw's core architecture and multi-agent deployment: bootstrap files, per-session transient agents, memory compaction, and sessions_send vs sessions_spawn.
DeepSeek Harness Hands-On: What the Other Half, Beyond the Model, Actually Buys You
Hands-on with DeepSeek Harness after its open-source release: all four entry points, reconstructing Trajectory from logs, and a comparison against Kimi Code.
Inside DeepSeek Harness: An Agent Architecture Built Entirely from Plugins
A teardown of DeepSeek Harness's plugin runtime: Fiber lifecycle, the two-layer scope chain, worker_threads isolation, and tool-shadowing — compared against Codex.
Taming AI Coding: A Team Playbook for Harness Engineering
The six pillars of Harness Engineering, MCP/Skills/knowledge-base setups in CodeBuddy, a three-phase rollout roadmap, and a harness-audit skill for team AI governance.
Is Loop Engineering Dead? A Guide to Graph Engineering
From Loop to Graph Engineering: the five-layer evolution, Loop's five flaws, the graph model, orchestration topologies, and cases from LangGraph, Uber, LinkedIn.
One person, one team: a minimal dev workflow with OpenSpec + Superpowers
OpenSpec locks in design intent, Superpowers runs TDD execution. One architect ships a tested kanban system from scratch in 30 minutes — full commands included.
Superpowers + gstack in Practice: 2 Plugins, 37 Skills, 5 Handoff Points, and a Complete Dev Loop
How Superpowers and gstack complement each other: Superpowers enforces code quality, gstack owns the product lifecycle — 5 handoff points, two integration patterns.
GStack in Practice: How AI Tooling Built This Blog From Scratch
GStack is a Claude Code skill system covering the full dev lifecycle, from product design to code review. Here's how I used it to build this blog in half a day.
Superpowers in Action: How AI Coding Turned Two Weeks of Work Into Half a Day
Using Claude Code to go from a vague idea to a fully deployed blog — design, review, testing, deployment — in half a day, versus two weeks without AI.