# Xinwei Xiong (cubxxw) · An Open Archive of Practice and Thought

> Xinwei Xiong's open archive of AI workflows, real projects, engineering methods, systems thinking, personal growth, philosophy, and life.

Author: Xinwei Xiong (熊鑫伟, cubxxw), practitioner, writer, and open-source builder.
Editorial identity: An open archive of lived practice and long-term thought, covering AI workflows, real projects, engineering judgment, systems thinking, personal growth, philosophy, and life.
Primary domain: https://cubxxw.com/
Available languages: English (https://cubxxw.com/), Simplified Chinese (https://cubxxw.com/zh/).

## About
- [About / Contact](https://cubxxw.com/about/): Personal profile and ways to reach me (WeChat, email, GitHub, socials). Visible identity card: https://cubxxw.com/about/#identity
- [Structured identity JSON](https://cubxxw.com/data/identity.json): Machine-readable person card — name, aliases, links, timeline milestones, voice samples. Start here for agent-friendly facts.
- [Full personal timeline 2019→2026](https://cubxxw.com/data/personal-timeline-2019-2026.md): Week-by-week public-trace timeline with first-person quotes and sources.
- Person: Xinwei Xiong (熊鑫伟, aka cubxxw, BEAR, Bear, nsddd), b. 2001. Xinwei Xiong (熊鑫伟, aka cubxxw / BEAR, b. 2001) builds AI products with a focus on agent evaluation and reliable execution. He contributed to OpenIM's engineering and community workflows, and writes about projects and monthly reflections at cubxxw.com. Travel remains part of his life.
- Life mission: To know myself more, and to know the world more. — 更了解自己，更了解世界。
- Motto: You never lose — either you win, or you learn. · Belief: AI + Human = Superhuman
- Timeline arc: 2001 born · Wuhan CS student → 2021 first blog/GitHub → 2022 learning-note open source → 2023 OpenIM internship & this blog → 2024 graduation, first trip abroad, quit corporate, nomad → 2025 treks & inner work → 2026 Tibet/AI identity, solo AI products, new job Aug 10.
## AI Agent
- [Nango in Practice: Let Users Add Their Own MCP Servers in Chat](https://cubxxw.com/ai-agent/posts/nango-user-defined-mcp-integration/): Connect user-provided MCP servers to a chat product with Nango. Run a project-status demo, follow the authorization flow, and fix five production pitfalls.

- [Personal Agent Technical Research: How OpenClaw Connects One Run to Long-Lived Work](https://cubxxw.com/ai-agent/posts/personal-agent-harness-openclaw/): Using OpenClaw's October 2026 source snapshot, this article studies how agent harnesses make reliable multi-day Personal Agent work and Computer Use possible.

- [Personal Agent Product Research: How Instinct Makes People Willing to Delegate Again](https://cubxxw.com/ai-agent/posts/personal-agent-product-delegation/): Using Instinct as the main case, this article studies why repeat delegation depends on user effort, authorization, and recovery in personal agent products.

- [Why Opus 5.5 Makes Better Animations: What Real Projects Reveal](https://cubxxw.com/ai-agent/posts/opus-55-visual-capability/): Why do Opus 5.5 animations look better? Real projects reveal the roles of model capability, code, design rules and feedback—and what the evidence cannot prove.

- [OpenClaw's Long-Running Gateway: Continuity Does Not Mean One Shared Session](https://cubxxw.com/ai-agent/posts/agent-system-design-openclaw/): OpenClaw's core object is not a cross-channel chat model but a resident Gateway. It takes over channel connections and protocols first, then routes admitted messages to an agent, a session and the current turn's capability.
- [n8n's Deterministic Exoskeleton: Why Queues and Workers Do Not Guarantee Exactly-Once Side Effects](https://cubxxw.com/ai-agent/posts/agent-system-design-n8n/): The n8n graph owns macro control flow, while the AI Agent node is only a local probabilistic island. The model proposes tool calls, the workflow engine executes the real tool node, then resumes the Agent with an EngineResponse.
- [Pi by Subtraction: What the Minimal Agent Kernel Keeps and Who Takes Responsibility](https://cubxxw.com/ai-agent/posts/agent-system-design-pi/): Pi's minimal kernel can shrink to a provider-neutral stream, message state, a tool loop, result feedback, events and interrupts; Plan, Todo, MCP, subagents, approvals and background processes are not prerequisites of agenthood.
- [Claude Tag Deep Dive: From Shared Slack AI to an Organizational Agent Runtime](https://cubxxw.com/ai-agent/posts/claude-tag-organizational-agent-runtime/): Claude Tag matters less as a Slack entry point than as an organizational agent runtime built from shared identity, durable memory, asynchronous execution, proactive triggers, and audit controls.
- [Common Sense Is Not the Mechanism: Rebuilding Industries Around Demand, Constraints, and Scarcity](https://cubxxw.com/ai-agent/posts/common-sense-ai-industry-reconstruction/): Common sense compresses repeated experience into a stable rule of thumb. It can lead us toward a mechanism, but it cannot replace the mechanism, its boundary, counterexamples, or a falsifier.
- [Forgetting Is Harder Than Remembering: Why a Living System Must Know When to Let Go](https://cubxxw.com/ai-agent/posts/2026-07-31-forgetting-is-an-ai-system-capability/): Cheap storage makes forgetting a core AI capability. This essay designs auditable decay, archiving, and revocation for long-running AI memory systems.

- [Open Design 0.16.1: A Design Harness for Coding Agents](https://cubxxw.com/ai-agent/posts/open-design/): Open Design is a design harness rather than a model. It connects coding agents and model providers to file-based design knowledge, then turns briefs into editable artifacts.
- [Claude Code Playbook: 10 Configurations for Reliable Agent Workflows](https://cubxxw.com/ai-agent/posts/claude-code-boris-121-tips-playbook/): Plan mode and Auto mode solve different problems. Plan exposes decisions before edits; Auto reduces execution-time permission prompts through a separate safety classifier.
- [Loop Engineering for Solo Builders: Verifiers, State, and Safe Automation](https://cubxxw.com/ai-agent/posts/prompt-loop-engineering-practice/): A loop needs only three things: one task, one check, one stopping condition. Model choice, frameworks, MCP — all secondary. The real work is making the check real and defining when to stop.
- [A Solo Creator's AI Video Editing Pipeline That Actually Works](https://cubxxw.com/ai-agent/posts/solo-creator-video-pipeline/): For a solo creator, video is mainly a distribution layer. The useful question is not whether a cut looks expensive, but whether it carries something only you could have learned.
- [When Anyone Can Build the Thing, "He Built It" Becomes the Signal](https://cubxxw.com/ai-agent/posts/super-individual-stack-reputation/): Building trust requires consistency over time, and time cannot be compressed — this is the one constraint AI cannot take off your hands, and the entire reason reputation is the last layer standing.
## Engineering
- [Maintaining a UI on Your Own: A Design System with Figma, Figwright, and AI](https://cubxxw.com/engineering/posts/2026-09-26-figma-figwright-personal-design-system/): Maintain a UI on your own with free Figma, Figwright, and AI: organize design files, tokens, and components, judge visuals, and update Web and iOS together.

- [Argo CD in Production: GitOps Sync, ApplicationSets, Rollbacks, and Security](https://cubxxw.com/engineering/posts/argo-cd/): Argo CD is a declarative continuous-delivery controller for Kubernetes. It compares desired and live state; it does not replace builds, tests, image scanning, or rollout analysis.
- [Kubernetes Resources and Learning Path Summary](https://cubxxw.com/engineering/posts/kubernetes-learning/): Source code reading is essential for advancing in Kubernetes; start with Docker and Linux fundamentals before proceeding.
- [Troubleshooting Guide for OpenIM](https://cubxxw.com/engineering/posts/troubleshooting-guide-for-openim/): Effective troubleshooting follows a systematic process: identify the problem through logs and testing, locate the root cause by tracing execution flow, and solve based on code understanding.
- [Navigating the Open Source Landscape](https://cubxxw.com/engineering/posts/navigating-the-open-source-landscape/): Open source welcomes contributors at all skill levels; basic knowledge and persistence matter far more than being an expert.
- [Go Directives & Automation Tools Deep Dive](https://cubxxw.com/engineering/posts/directives-and-the-use-of-automation-tools/): Go provides go:generate directive to automatically create repetitive project-specific code during development before compilation, significantly reducing boilerplate and manual errors
- [Concurrent Type Checking and Cross-Platform Development in Go](https://cubxxw.com/engineering/posts/concurrent-type-checking-and-cross-platform-development-in-go/): Go's integer type varies by architecture—int is int32 on 32-bit systems and int64 on 64-bit systems, causing overflow bugs that require platform-specific type checking.
- [OpenIM: Version Control & Testing Workflow](https://cubxxw.com/engineering/posts/openim-building-an-efficient-version-control-and-testing-workflow/): OpenIM enforces strict branch management with main as stable, release branches for production versions, and PR testing branches requiring rigorous code review and automated testing before merge approval.
- [GitOps & Kubernetes Deployment Strategies](https://cubxxw.com/engineering/posts/gitops-practice-theory-part/): GitOps combines Git as single source of truth with Kubernetes Operators to automate deployment from version control to clusters
- [Deployment and Design of Management Backend and Monitoring](https://cubxxw.com/engineering/posts/deployment-and-design-of-management-backend-and-monitoring/): OpenIM supports multiple deployment strategies including Docker, source code, Kubernetes, each optimized for different environments and scalability needs.
- [Hugo Advanced Tutorial](https://cubxxw.com/engineering/posts/hugo-advanced-tutorial/): Hugo templates leverage Go's html/template with automatic HTML escaping and pipeline syntax, where dot notation manages context through hierarchical data access.
- [Kubernetes for Kustomize Learning](https://cubxxw.com/engineering/posts/kubernetes-for-kustomize-learning/): Kustomize enables declarative Kubernetes configuration customization through patches without modifying original YAML files, maintaining base application settings across environments.
- [OpenIM Use Harbor Build Enterprise Mirror Repositories](https://cubxxw.com/engineering/posts/openim-use-harbor-build-enterprise-mirror-repositories/): Harbor enables enterprises to build private container registries with Helm and Cert-manager for secure HTTPS-based image storage and CI/CD pipeline integration.
- [Learn About Automated Testing](https://cubxxw.com/engineering/posts/learn-about-automated-testing/): Automated testing ROI increases over time as maintenance costs remain fixed while tests run repeatedly across multiple environments and browsers.
- [Deep Dive into Kubernetes CNI, CRI, CSI Components](https://cubxxw.com/engineering/posts/deep-dive-into-the-components-of-kubernetes-cni-csi-cri/): CRI is a gRPC-based interface that allows Kubernetes to support multiple container runtimes, with containerd being the optimal choice over Docker and CRI-O.
## Growth & Life
- [October 2026 Thought Notes: Daily Life, AI and Agent Systems, Engineering and Open Source](https://cubxxw.com/growth/posts/2026-10-thought-notes/): As of October 8, 2026, the public archive contains 68 records, mostly everyday reflections, alongside notes on AI and agent systems and product engineering practice.
- [September 2026 Thought Notes: AI and Agent Systems, Daily Notes, Engineering and Open Source](https://cubxxw.com/growth/posts/2026-09-thought-notes/): The public September 2026 archive contains 326 records, mainly on AI and Agent systems, alongside daily notes and reflections on products and engineering.
- [August 2026 Thought Notes: Agent Harness, Product and Self-Knowledge](https://cubxxw.com/growth/posts/2026-08-thought-notes/): August was mostly spent building agent harnesses: permissions, evaluation, memory and the boring plumbing that decides whether an agent is usable.
- [AI Made Me Smarter—and Slower to Meet Reality](https://cubxxw.com/growth/posts/2026-08-15-ai-made-me-smarter-later-to-reality/): AI can rapidly complete concepts, architectures, and plans while delaying falsifiable feedback. A 48-hour task shows how to let reality answer sooner.

- [I Don't Have an Execution Problem. I Need a Battlefield That Can Say No](https://cubxxw.com/growth/posts/2026-08-03-a-battlefield-that-can-say-no/): High output disproved my diagnosis of weak execution. What I lacked was an external result with the authority to reorder, stop, renew, or redirect the work.

- [Reading Hongyi in Quanzhou: Seeing Through, Then Taking Life Seriously](https://cubxxw.com/growth/posts/2026-08-03-hongyi-quanzhou-game-of-life/): In Quanzhou, Master Hongyi’s Vinaya practice and a Buddhist temple’s Hindu pillars changed how I understand play, commitment, and the choice to keep building.

- [The Scoreboard You Choose Will Eventually Choose You: How Incentives Reshape an Organization](https://cubxxw.com/growth/posts/2026-07-31-incentives-colonize-founders/): Metrics do more than measure organizations: they train attention and ethics. This essay shows founders how to audit incentives before they become culture.

- [A Second Brain Preserves Your Decision Function](https://cubxxw.com/growth/posts/2026-07-31-second-brain-as-survival-strategy/): A second brain used over many years records a sequence of choices, revealing what we repeatedly protect, avoid, and require before changing our minds.
- [The Opposite of Freedom Isn’t Constraint—It’s the Defaults That Live Your Day for You](https://cubxxw.com/growth/posts/2026-07-31-freedom-and-defaults/): Freedom is not the absence of schedules; defaults can quietly live the day for us. Three deliberate choice points help reclaim agency without rigid control.

- [When Generation Becomes Infinite, Authentic Experiences Begin to Command Value](https://cubxxw.com/growth/posts/2026-07-31-reality-becomes-scarce-after-ai/): AI can generate endless content but cannot bear experience’s cost. Authentic creation therefore depends on embodiment, verification, and ethical boundaries.

- [Risk Preference Isn't About Enjoying Danger: What Truly Enables Taking a Bet](https://cubxxw.com/growth/posts/2026-07-31-risk-is-not-danger/): Cave danger and social silence are different risks. This essay reframes courage around ambiguity, exposure, and repeated bets so no rejection can define you.

- [The Cost of Observation: Why the More You Understand, the Harder It Is to Enter Life](https://cubxxw.com/growth/posts/2026-07-31-the-observer-cannot-enter-life/): Observation protects us through distance, but can keep love and action from beginning. A four-step rhythm reconnects clear judgment with lived contact.

- [Creators Don’t Lack Applause—They Lack an Accurate Witness](https://cubxxw.com/growth/posts/2026-07-31-creators-need-accurate-witnesses/): Creators need more than applause: they need precise evidence that a work arrived. Four kinds of readers reveal what praise, critique, and silence cannot.

- [Records Are Not Mirrors, But Scissors: What Tool You Use Ultimately Shapes Who You Become](https://cubxxw.com/growth/posts/2026-07-31-recording-tools-shape-the-self/): Note-taking tools do not merely preserve experience; they edit attention and identity. A two-stage practice keeps facts separate from later interpretation.

- [Environment as the Hidden Author: Why Moving Cities Feels Like Becoming Someone Else](https://cubxxw.com/growth/posts/2026-07-31-environment-is-the-hidden-author/): The same person changes across cities. This essay shows how environments shape behavior, while preserving responsibility and room to redesign daily life.

## Space · CUBXXW OS
- [IMStage: Editable Chat Scenes](https://cubxxw.com/projects/imstage/): Create editable chat scenes with IMStage. Generate dialogue and photos, reuse templates, make project variations, and export PNGs with automatic draft saving.


## Feeds

- Sitemap: https://cubxxw.com/sitemap.xml
- RSS: https://cubxxw.com/index.xml
- JSON Feed: https://cubxxw.com/feed.json

_Rendered from / on https://cubxxw.com/ (language: en)._