PROJECT REGISTER / 002
Start with the problem and key judgement, then follow the prototype, current evidence, and next step.
Search infrastructure that agents can inspect and control
The decision was not to find a cheaper search API, but to treat search as a product capability with multi-source verification, progressive disclosure, and replaceable engines.
An external-PR investigation that scanned 93 open-source agent projects selected it as an architecture-validation case in the MCP search category.
Reversible MCP context compression
Reduce the context an agent receives first while preserving selected arguments, one upstream call, and exact recovery.
Across a frozen set of 12 tools and 24 bilingual tasks, normal-path tokens fell from 71,388 to 17,007; all 24 tasks completed with 24 upstream calls.
Keep AI coding agents inside the current task boundary
Do not judge overengineering by line count or keywords. Check user authority, current consumers, reachable evidence, and whether omission would fail current acceptance.
The public 0.2.0 release provides adapters for Codex, Claude Code, OpenCode, Hermes Agent CLI, and Pi, adds the standalone Stop That Shit Slop (STSS) Skill, and keeps paired Bad/Good Cases that preserve necessary work.
Make AI collaboration persist beyond a single session
Treat documentation, handover, and reusable skills as product information architecture rather than post-delivery paperwork.
The practice record covers trajectory mining across 196 sessions and turns the learning into reusable skills and a maintenance loop.
OPEN-SOURCE INDEX / 005
These projects do not need full case files yet, but together they show my work across product workflows, agent tooling, and collaboration infrastructure.
Product workflow
An AI workflow for Chinese support tickets, combining hierarchical retrieval, evidence-backed drafting, human review records, and offline evaluation.
Python↗06Content system
A 12-Factor content system for AI agents, spanning multi-source extraction, SEO generation, and knowledge evolution.
Python↗07Agent tooling
An SEO crawling and auditing tool built for AI agents, with both CLI and MCP interfaces.
Python↗08Collaboration infrastructure
A reference architecture for multi-project AI workspaces, organising AGENTS, HANDOVER, and lifecycle hooks into a portable system.
Shell↗09Orchestration
A reusable AI orchestration system for complex multi-stage development work, with explicit phase boundaries and handoffs.
Shell↗10Knowledge tooling
An Open Knowledge Format navigator for searching, auditing, fixing, and exporting knowledge bundles.
Python↗11Skill system
A Claude Code skill collection for frontend and Next.js workflows, combining curated community skills with custom practice.
Python↗12Content pipeline
A pluggable YouTube-to-article pipeline covering download, AI rewriting, SEO, and HTML rendering.
Python↗