Blog
Notes from operating an AI team in production — Harness Engineering, multi-agent design, execution chain governance.
Scaling AI Game Production with Multi-Agent Stability
Scaling AI Game Production with Multi-Agent Stability
Preventing Goal Drift in Multi-Agent Systems
Preventing Goal Drift in Multi-Agent Systems
Synapse CLI Activation: Whitelist and D-record Pitfalls
Synapse CLI Activation: Whitelist and D-record Pitfalls
GStack Integration Audit: Multi-Product Architecture Fragmentation
GStack Integration Audit: Multi-Product Architecture Fragmentation
Local-First Proxy: Solving Synapse Cache Sync Delays
Local-First Proxy: Solving Synapse Cache Sync Delays
Automating AI News Translation with Claude Code
Automating AI News Translation with Claude Code
Bridging Product Intent and System Behavior Through Verification
Bridging Product Intent and System Behavior Through Verification
GBrain's 8-Layer Memory Architecture: Lessons from Production
GBrain's 8-Layer Memory Architecture: Lessons from Production
L0-L5 Pyramid: Scaling Obsidian Knowledge Management
L0-L5 Pyramid: Scaling Obsidian Knowledge Management
Building a Team Knowledge Base: From Fragmentation to Structure
Building a Team Knowledge Base: From Fragmentation to Structure
Multi-Agent Pipeline Orchestration for Automated Meeting Processing
Multi-Agent Pipeline Orchestration for Automated Meeting Processing
Bilingual Strategy Reduces Claude Code Token Costs
Bilingual Strategy Reduces Claude Code Token Costs
Transforming Scattered Notes into AI-Ready Knowledge Systems
Transforming Scattered Notes into AI-Ready Knowledge Systems
AI Decision Drift After OBS Upgrade: A Postmortem
AI Decision Drift After OBS Upgrade: A Postmortem
Synapse Latency Debugging in Multi-Product Architecture
Synapse Latency Debugging in Multi-Product Architecture
Handling Business Days in Gantt Chart Date Calculations
Handling Business Days in Gantt Chart Date Calculations
Agent-Based Verification Cuts Crawling False Alarms
Agent-Based Verification Cuts Crawling False Alarms
Prompt as Code: Fixing AI Programming Strategy Gaps
Prompt as Code: Fixing AI Programming Strategy Gaps
Task Topology Design for Stateful Agent Pipelines
Task Topology Design for Stateful Agent Pipelines
Why L4-to-L5 Agent Upgrades Fail Without Value Frameworks
Why L4-to-L5 Agent Upgrades Fail Without Value Frameworks
Detecting Dormant n8n Workflows with REST API
Detecting Dormant n8n Workflows with REST API
Dual-Layer Tag Model for Cross-Domain Knowledge Retrieval
Dual-Layer Tag Model for Cross-Domain Knowledge Retrieval
Passive Storage to Active Indexing for Agent Knowledge Bases
Passive Storage to Active Indexing for Agent Knowledge Bases
Routing L1-L2 Decisions Back to Machines via Agent Memory Collaboration
How externalized memory with active_wbs fields reduced task recovery from 5 minutes to 30 seconds, routing L1-L2 decisions away from human approval bottlenecks.
Agent Memory Collaboration Mechanism Design
Agent Memory Collaboration Mechanism Design
Ai Blog Content Quality Degradation Root Cause Fix
How a structural_qa threshold misconfiguration lowered from 75 to 30 silently bypassed the quality gate for six weeks in the AI blog pipeline.
AI Platform Migration: Adapter Pattern for Seamless Switching
AI Platform Migration: Adapter Pattern for Seamless Switching
Blog Content Quality Regression Systematic Fix
structural_qa threshold dropped from 75 to 30 during a visual redesign, silently degrading blog content quality for 6 weeks with no errors raised.
Minimal Impact Multilingual Database Migration Notion
Dual-database coexistence strategy for migrating 1,200 Notion records from Chinese to English field names across 17 n8n nodes with zero downtime.
Multi-Agent Orchestration: Decoupling Execution from Decision Chains
Multi-Agent Orchestration: Decoupling Execution from Decision Chains
n8n Cross-Environment Migration: Credentials and Variables
n8n Cross-Environment Migration: Credentials and Variables
SSOT Pattern Solves Multilingual AI Workflow Maintenance
SSOT Pattern Solves Multilingual AI Workflow Maintenance
Cross-Border Lease Arbitrage: Why Simple Interest Rate Gaps Fail
Cross-Border Lease Arbitrage: Why Simple Interest Rate Gaps Fail
Self-Healing GitHub Actions Runners via Heartbeat Monitoring
Self-Healing GitHub Actions Runners via Heartbeat Monitoring
Multi-Agent Authorization in AI Workflows
Multi-Agent Authorization in AI Workflows
Multi-Agent Collaboration for Project Template Standardization
Multi-Agent Collaboration for Project Template Standardization
Real-Time Pipeline Health Management in PMO Auto v2.6.0
Real-Time Pipeline Health Management in PMO Auto v2.6.0
Validating Knowledge Tool ROI in Multi-Agent Systems
Validating Knowledge Tool ROI in Multi-Agent Systems
[EN] 如何在3-5人小团队内推广 AI 协作体系：从零依赖 Demo 开始
不从体系文档入手，而是用一个能跑的 Demo 让同事第一次感受到 AI 多 Agent 的价值
[EN] 让 AI 系统自我进化：用 audit log 构建行为约束的反馈闭环
CEO Guard 记录每次工具调用违规，周维度分析将异常模式转化为新规则，系统越用越守纪律
Designing Output Path Rules for AI Workflows: Why Quick Fixes Become System Debt
One misfiled PDF revealed 9 path errors over 47 days. Here's why prompt-based fixes don't solve structural problems, and how to move output routing logic from agent prompts to explicit config.
Chinese Token Optimization in Multi-Agent Systems
Chinese Token Optimization in Multi-Agent Systems
Synapse Three-Layer Prompt Architecture for Team Collaboration
Synapse Three-Layer Prompt Architecture for Team Collaboration
[EN] 当AI遇上多语言网站：一次演示前夜的系统性质量修复
以真实演示压力为背景，展示如何用 AI 团队在24小时内完成内容质量、数据准确性、多语言完整性三层审查并落地修复
How a 4-Agent Team Audited 47 Pages — and Why Authority Design Is the Hard Part
A live test of multi-agent website auditing: how we structured committee authority, role-based output schemas, and a dedicated opposition agent that blocked 4 SEO-damaging changes.
When Migration Costs More Than Rebuilding: A PMO Product Phoenix Story
At week 14, migration would take 27 weeks instead of 6. We stopped, discovered 40% of features existed only to manage the system's own complexity, and rebuilt from scratch.
Why We Scheduled All Product Pipeline Refreshes at 2AM
Stale data cost a client a pricing miss — and every pipeline log was green. How we moved from 'someone will remember' to an automated daily refresh with freshness timestamps exposed to decision-makers.
[EN] 产品管线的知识新鲜度问题：为什么每天2AM要跑一次全局更新
从「涉及产品线时需要重新整理信息」的痛点出发，阐述产品知识自动化更新机制的设计逻辑
From Solo AI to Agent Teams: How Lysander Leads a Virtual Team Through Autonomous Product Delivery
A real-world walkthrough of the AI organizational design behind Synapse: user authorization, Lysander coordination, specialist decision-making, and committee collaboration — and how it delivers autonomously.
Multilingual PMO Automation: Architecture-First i18n Strategy
Multilingual PMO Automation: Architecture-First i18n Strategy
AI Team Decision Boundaries: When Should You Let the Agent Decide?
From 'ask me every 5 minutes' to 'professional decisions by professionals' — how we evolved Claude Code delegation in a real engineering team
Why Your AI Knowledge Base Decays: Lessons from CLAUDE.md Fact Accuracy Failures
The moment I discovered all the agent count data was wrong — how config file decay makes every AI analysis rest on sand
Replacing Local Cron with n8n: A Complete Intelligence Pipeline Migration
From discovering overlapping scheduled task responsibilities to systematically cleaning redundancy — establishing a single source of truth after migration
When Your AI Team Has 3 Different Headcounts: Fixing Number Drift with Fact-SSOT
From a real incident where 44/46/50 coexisted as the 'correct' agent count — root cause analysis of number drift in multi-file AI systems and the Fact-SSOT meta-rule solution
Replacing a 15-Person Outsourced Team with AI: A Financial Leasing Company's Synapse Transition
From a real client scenario: how 'product manager + AI' can absorb work that previously required a 15-person outsourced development team
The AI Time Illusion: A System-Level Analysis of Cross-Day Conversation Failures
A 2am email listing Wednesday tasks as 'completed today' on Thursday — dissecting why long AI sessions systematically lose temporal grounding and three layers of engineering fix
Building WF-09: The n8n Slack Notification Hub Router — Architecture Deep Dive
7 workflows, each with its own Slack credentials and zero audit trail. The complete design and migration story for WF-09, a centralized n8n notification router.
The AI Time Illusion: Why Cross-Day Conversations Generate False 'Tonights'
Session continuity ≠ time continuity — how AI fails at temporal reasoning in long or resumed conversations, and how to fix it with explicit time anchoring
Fact-SSOT: The Meta-Rule for Eliminating Number Drift in AI Agent Systems
When the same AI system had agent counts of 44, 46, and 50 in three files simultaneously — root cause analysis and the Fact-SSOT meta-rule as a systematic fix
Unified Slack Notification Routing with n8n: From Fragmented Direct Connections to a Single Auditable Pipeline
7 workflows directly connecting to Slack — credential sprawl, format inconsistency, zero observability. The WF-09 hub router pattern that fixed it.
Notes from the Field: GA Validation and Version-Locking for AI Products
An abstract of a Chinese case study on the full GA acceptance flow for PMO Auto V2.0 — structured requirements pool, layered test suites, and the mandatory five-step version lock.
Notes from the Field: The Pipe Wasn't Broken, the Contract Was
An abstract of a Chinese root-cause analysis on a silent Slack notification failure caused by payload contract drift between an Agent caller and an n8n Webhook receiver.
Designing an AI CEO That Cannot Skip the Process
From the CEO Guard audit hook to mandatory dispatch tables to P0 violation logs — the governance architecture that puts an AI executive on rails.
Notes from the Field: Two Months of n8n + Claude PMO Automation Lessons
An abstract of a Chinese retrospective on building a PMO automation system with n8n and Claude — credentials hygiene, Notion rate-limit traps, and why state-based E2E acceptance criteria matter.
How to Keep a Multi-Agent System From Quietly Rotting
A working evolution framework for multi-agent teams: capability-card audits, harness rule-entropy controls, and an intelligence pipeline that actually changes the system.
Notes from the Field: When Your AI Quietly Broadcasts to the Whole Company
An abstract of a Chinese incident report on discovering an AI scheduled task had been posting daily intel digests to #general (287 people) for three weeks, and the four governance rules we instituted afterward.
Notes from the Field: When to Stop Debugging and Switch to MCP
An abstract of a Chinese case study on abandoning a Notion API script after hitting permission walls, and using Claude's MCP integration as a strategic re-architecture rather than a workaround.
Training Vertical AI Agents With Product Test Cases: Making an Agent Truly 'Get' a Product
Test cases aren't just QA tools — they're the best vehicle for systematically accumulating product-domain knowledge in AI Agents.
Notes from the Field: Phantom Data in Automation Pipelines
An abstract of a Chinese case study where 111 records were written to Asana but only 19 were visible to the gate check, exposing a structural data-consistency trap in multi-stage pipelines.
An AI CEO's First 'Autonomous Decision Authority': When the User Says 'You Decide, Don't Ask Me'
Starting from one sentence — 'Let Lysander CEO organize the work, not me' — exploring how authorization boundaries evolve in human-AI collaboration.
How One Missing git init Brought Down an AI Team for 4 Days
A real-incident retro showing how a single point of failure in the infrastructure layer can cascade through an AI automation system.
Notes from the Field: Time-Window Exemptions for AI Guard Rails
An abstract of a Chinese case study on resolving the conflict between CEO Guard (a high-risk-tool blocker) and unattended scheduled agents, by exempting guard rules within pre-registered time windows.
When Automation Fails Silently 23 Times: Why Zero Logs Are More Dangerous Than Errors
23 consecutive zero-message triggers from pmo-wbs-trigger exposed a counterintuitive truth: 'fake success' is harder to debug than 'real errors.'
Notes from the Field: When Scheduled Tasks Run 19 Times and Produce Nothing
An abstract of a Chinese case study on diagnosing a silent-failure n8n trigger that fired 19 times with green status but zero downstream output, and the three-layer guard rails we added afterward.
Notes from the Field: Merging Four Drifted Copies of My AI System
An abstract of a Chinese case study on consolidating four parallel Synapse instances (Mini, Work, Lab, Archive) that had drifted apart over two months, and the dimension-by-dimension merge strategy used.
Notes from the Field: An AI Team's First Video Production Pipeline
An abstract of a Chinese case study on validating a brand-new capability (video production) by mapping the chain end-to-end, finding the weakest link first, and designing a multi-role script review.
Synapse Agents Market: Designing a Distributable Product From an Internal AI Collaboration System
Borrowing from the online-course platform model to package a private AI team system as an external product, with decision logic for a dual-track validation strategy.
From a 44-Agent Monolith to a Modular Agent Marketplace
A real architectural evolution: how a multi-agent team broke free of monolithic, department-shaped configuration and became a set of installable, composable capability modules.
After an AI Agent Deleted My Files: A Painful Retro on Execution Boundaries
A real incident as the entry point for analyzing where the safety boundary should sit between AI autonomous execution and human confirmation.
Notes from the Field: Personal Task Management with Claude Code as the Hub
An abstract of a Chinese article on collapsing capture, organize, and execute into a single Claude Code session, using a YAML active_tasks file as the only state store.
Building an AI Company With an Actual Execution Chain
Not 'AI helps me do work' — 'an AI company that operates.' The design notes behind Synapse: CEO role, decision chain, dispatch protocol, and QA gates.
Notes from the Field: AI Tools Talk — A Tour of My Personal AI Stack
An abstract of a Chinese presentation deck covering cc-connect WeChat integration, Obsidian second brain, Harness Engineering, the 29-agent Synapse team, and a stock trading system.
Notes from the Field: How to Build a Second Brain with Claude Code (Slides)
An abstract of a Chinese presentation deck on building a Software 2.0 second brain that auto-extracts decisions, concepts, problems, learnings, and projects from Claude Code conversations.
Don't Rush to Install gstack — First Think About Who Should Use It
YC President's open-source gstack lets one person become a team. The higher-leverage move is to take it apart and graft it onto your existing team's DNA.
Claude Code as a Zero-Friction Deployment Engine: Bringing Non-Technical Colleagues Onto the AI Team
How Claude Code's automatic dependency handling removes the 'last-mile' friction that blocks AI tool rollout to non-engineers.
Copy Your AI Collaboration System in One Prompt: How to Hand Synapse to a Colleague
From 'I built an AI collaboration system' to 'a colleague reproduces it with one prompt' — the practice of Prompt-as-deployment-doc.
Notes from the Field: When Principles Fail, Encode Them in Code
An abstract of a Chinese article on transforming team decision principles from documentation into code-level enforcement, eliminating the 'everyone knows the rule, everyone breaks the rule' anti-pattern.
Notes from the Field: Obsidian as the Single Source of Truth for an AI Team
An abstract of a Chinese article on using Obsidian as the SSOT for an AI multi-agent team — HR knowledge, decision rules, and Harness Engineering all derive from Obsidian markdown cards.
Notes from the Field: Building a Second Brain from Claude Code Conversations
An abstract of a Chinese article on automatically extracting decisions, concepts, problems, learnings, and project notes from Claude Code .jsonl conversation logs into an Obsidian knowledge base.
Notes from the Field: Claude Code Self-Healing Error System
An abstract of a Chinese deep-dive on building a confidence-driven error analysis pipeline for Claude Code using n8n, a Python middleware, and Slack notifications.
Harness Engineering for AI Agents: Why Your Model Isn't the Bottleneck
A practitioner's guide to Harness Engineering — the discipline of building feedforward and feedback controls so your coding agents stop repeating the same mistakes.
Notes from the Field: An n8n Workflow That Publishes to WeChat Drafts
An abstract of a Chinese implementation log for an n8n workflow that scrapes Astro/Tailwind blog HTML and pushes a stripped-down version to WeChat Official Account drafts.
Notes from the Field: Understanding n8n's executeOnce Setting
An abstract of a Chinese tutorial on n8n's item-based execution model and the executeOnce flag, with examples for chained Asana/Slack workflows.
Notes from the Field: An Asana + Slack + n8n PMO Pipeline
An abstract of a Chinese case study describing a project-management automation that auto-generates task chains from a CSV process table and notifies the right person on completion.