https://blog.chuanxilu.net/en/Recent content on Chuanxilu for Skilled Homo sapiensHugoen-USFri, 22 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/dual-pass-review-recall-precision-tradeoff/Fri, 22 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/dual-pass-review-recall-precision-tradeoff/A design review agent needs to find every issue AND avoid false positives. One agent can't do both well. Borrowing cascade retrieval from information retrieval — a 15-year-old method — I split the agent into two: one for recall, one for precision. Real defects get caught earlier, and the risk of rework during development drops.https://blog.chuanxilu.net/en/posts/2026/05/ai-practice-15-turn-conversation/Fri, 22 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-practice-15-turn-conversation/Series companion practice: pick a multi-step task and hold at least 15 turns of conversation with AI. Experience drift and recovery in long conversations, and practice progress summaries and context management.https://blog.chuanxilu.net/en/posts/2026/05/tdd-pipeline-phase7-invisible-gap/Thu, 21 May 2026 11:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/tdd-pipeline-phase7-invisible-gap/Phase 6 drills every bug to root cause. But it doesn't scan for shared patterns across bugs, unchecked component gaps, or execution order flaws. That's Phase 7's job. In small systems it catches more bugs. In large systems, the same findings point to architectural evolution. Phase 7 doesn't make architecture decisions — it provides the scarcest input for them: evidence-based problem localization.https://blog.chuanxilu.net/en/posts/2026/05/ai-path-l0-l1-week3/Thu, 21 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-path-l0-l1-week3/Part 3 of the AI Path L0→L1 Upgrade Guide: follow-up iteration is the most underrated skill in multi-turn conversation, context management keeps AI from drifting off track, and role-playing unlocks entirely different depths of insight from the same question.https://blog.chuanxilu.net/en/posts/2026/05/tdd-pipeline-v07-refinement-experiment/Wed, 20 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/tdd-pipeline-v07-refinement-experiment/I built a ruler. The ruler measured 'redundancy is harmful.' Then I used that ruler to trim the ruler's own redundancy. I deleted the operational steps from my AI skill files, keeping only principles and counterexamples. The model reconstructed the deleted steps on its own — output quality didn't drop.https://blog.chuanxilu.net/en/posts/2026/05/ai-tip-format-constraints/Wed, 20 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-tip-format-constraints/Tired of reformatting AI responses every time? Six common format constraints, each with a ready-to-use prompt template you can copy and paste directly into your query.https://blog.chuanxilu.net/en/posts/2026/05/tdd-pipeline-v08-failed-experiment-discovery/Tue, 19 May 2026 18:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/tdd-pipeline-v08-failed-experiment-discovery/I tried refining the pre-release testing phase of my TDD Pipeline by replacing step-by-step instructions with principles. The refined version failed at its core job. But comparing where it failed against where it unexpectedly succeeded revealed that individual defect diagnosis alone wasn't enough — it needed a systematic scanning layer on top.https://blog.chuanxilu.net/en/posts/2026/05/ai-tip-when-to-use-cot/Tue, 19 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-tip-when-to-use-cot/Chain-of-Thought can dramatically improve AI output quality — but you shouldn't use it every time. Three signals to help you decide: when adding 'please reason step by step' helps, and when it just wastes time.https://blog.chuanxilu.net/en/posts/2026/05/ai-5-rbgo-examples/Mon, 18 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-5-rbgo-examples/The RBGO framework sounds straightforward, but when you actually sit down to write a prompt, it's easy to get stuck. Here are 5 everyday scenarios—each with a full side-by-side comparison of the vague version and the rewritten version. Emails, analysis, learning, planning, code review. Copy them, use them directly.https://blog.chuanxilu.net/en/posts/2026/05/why-articulation-upgrade-and-takeaways/Sun, 17 May 2026 09:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/why-articulation-upgrade-and-takeaways/Upgrading the Why Articulation template based on A/B test data: replacing explicit questions with open-ended reasoning plus self-check, keeping mandatory tone and negative-only examples. Three transferable prompt engineering lessons.https://blog.chuanxilu.net/en/posts/2026/05/ai-practice-rbgo-rewrite/Sun, 17 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-practice-rbgo-rewrite/Series practice challenge: rewrite your first AI question of the day using the RBGO framework. Same need, four extra lines of context — see how dramatically the answer improves.https://blog.chuanxilu.net/en/posts/2026/05/ai-path-l0-l1-week2/Sat, 16 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-path-l0-l1-week2/Part 2 of the AI Path L0→L1 Upgrade Guide. Master the RBGO prompt framework (Role–Background–Goal–Output), learn Chain-of-Thought reasoning to improve analytical answers, and use format constraints to make AI output ready to use.https://blog.chuanxilu.net/en/posts/2026/05/ab-test-positive-examples-harm/Fri, 15 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ab-test-positive-examples-harm/Why do positive examples make AI output worse? A 4-variable A/B test on Why Articulation structure, tone, position, and example type found that demonstrations hurt — echoing Anthropic's alignment research.https://blog.chuanxilu.net/en/posts/2026/05/ai-tip-which-ai-to-use/Fri, 15 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-tip-which-ai-to-use/Tried ChatGPT, Claude, Gemini — still not sure which one to use? A scenario-based framework to find the right fit.https://blog.chuanxilu.net/en/posts/2026/05/anthropic-alignment-to-prompt-design/Thu, 14 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/anthropic-alignment-to-prompt-design/Anthropic discovered that teaching models "why" works better than teaching them "what" — misalignment dropped from 22% to 3%. This insight from safety training applies to everyday prompt design too.https://blog.chuanxilu.net/en/posts/2026/05/ai-tip-working-vs-long-term-memory/Thu, 14 May 2026 06:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-tip-working-vs-long-term-memory/Why your AI forgets what you said five minutes ago — and what to do about it.https://blog.chuanxilu.net/en/posts/2026/05/ai-3-cognitive-shifts/Wed, 13 May 2026 00:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-3-cognitive-shifts/Open ChatGPT, type a keyword, copy the answer, close the tab. That works fine — but it wastes 90% of what AI can do. Three real scenarios show you what changes when you shift how you think about AI.https://blog.chuanxilu.net/en/posts/2026/05/ai-practice-same-question-3-times/Tue, 12 May 2026 08:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-practice-same-question-3-times/Series practice challenge: ask an AI the same open-ended question three separate times and compare the answers. Experience the probabilistic nature of LLM generation firsthand—and stop treating AI like a search engine.https://blog.chuanxilu.net/en/posts/2026/05/ai-path-l0-l1-week1/Mon, 11 May 2026 08:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-path-l0-l1-week1/First part of the AI Path L0→L1 Upgrade Guide series. LLMs aren't search engines—they generate answers rather than retrieve them. Understand the difference between working memory and long-term memory, learn the strengths of mainstream platforms, and build the cognitive foundation for the next 4 weeks of practice.https://blog.chuanxilu.net/en/posts/2026/05/ai-toolchain-evolution-path/Sun, 10 May 2026 08:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-toolchain-evolution-path/An evolution map of AI capabilities from L0 to L4—not a tutorial for any specific tool, but a guide to understanding the fundamental mindset shifts at each stage. Includes an interactive HTML page where you can explore detailed skill checklists, recommended tools, and transition conditions for each level.https://blog.chuanxilu.net/en/posts/2026/05/design-doc-management-lessons-from-three-projects/Fri, 08 May 2026 15:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/design-doc-management-lessons-from-three-projects/AI-assisted development generates tons of design documents that live in .gitignore, invisible to git. A single rebase silently deletes them, and git reflog can't bring them back. This post walks through a lightweight git worktree setup that protects these documents, backed by real project data.https://blog.chuanxilu.net/en/posts/2026/05/six-bug-patterns-and-integration-gaps/Thu, 07 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/six-bug-patterns-and-integration-gaps/Before releasing Aristotle v1.1, I found 18 bugs. Unit tests caught four. The rest lived at the integration layer. After root cause analysis, six patterns emerged — not because the problems got harder, but because AI bypassed the defenses I'd built through years of experience.https://blog.chuanxilu.net/en/posts/2026/05/omo-vs-omo-slim-token-comparison/Wed, 06 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/omo-vs-omo-slim-token-comparison/I switched from OMO to SLIM expecting lower token bills. The total average barely moved. But when I broke it down by task type, the picture got far more interesting.https://blog.chuanxilu.net/en/posts/2026/05/inquiry-protocol-design-3/Wed, 06 May 2026 09:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/inquiry-protocol-design-3/The inquiry protocol's last line of defense: an independent confirmer uses falsifiability testing to hunt for counterexamples, pushing back against AI's shallow anchoring. Plus the protocol's origin story — from 18 bugs of practice to the v0.4.1 upgrade, and plans for reflexivity.https://blog.chuanxilu.net/en/posts/2026/05/inquiry-protocol-design-2/Tue, 05 May 2026 17:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/inquiry-protocol-design-2/Designing termination conditions for an inquiry protocol: T1–T3 are floor conditions (ensure AI goes deep enough), HC1–HC4 are guardrails (prevent the inquiry from spiraling). T2's preventive counterfactual check is the most important insight.https://blog.chuanxilu.net/en/posts/2026/05/inquiry-protocol-design-1/Tue, 05 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/inquiry-protocol-design-1/5-Why handed to AI fails not because the method is outdated, but because AI thinks shallow—stopping early, chasing one thread, seeking only confirming evidence. A real case where all four rounds of attribution went wrong.https://blog.chuanxilu.net/en/posts/2026/05/ai-bug-root-cause-diagnosis/Fri, 01 May 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/05/ai-bug-root-cause-diagnosis/Fix #13, and old bugs come back. This isn't a "how I fixed a bug with AI" anecdote. It's a full post-mortem of a 15+ bug battle — four rounds of attribution, regression traps, and how TDD was forced out by pain.https://blog.chuanxilu.net/en/posts/2026/04/ai-tdd-full-pipeline-from-requirements-to-code/Thu, 30 Apr 2026 14:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/ai-tdd-full-pipeline-from-requirements-to-code/Article 6 in the series. The previous five articles each covered one layer — requirements, design, testing, review, procedural justice. This one connects them into a working pipeline. Checklists for every stage, a real-project retrospective, and a blunt assessment of when this process is worth the overhead.https://blog.chuanxilu.net/en/posts/2026/04/adversarial-review-critical-thinking-ai-quality/Thu, 30 Apr 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/adversarial-review-critical-thinking-ai-quality/Ralph Loop v0.3 encodes procedural justice into the review protocol — structured review output, critical scrutiny, contested issue protocol — so every review decision has evidence, records, and rule-based constraints. Inspired by Robert's Rules of Order, born 150 years ago.https://blog.chuanxilu.net/en/posts/2026/04/ralph-loop-ai-errors-converge/Wed, 29 Apr 2026 14:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/ralph-loop-ai-errors-converge/The tech spec from the previous article shrank the space where AI can improvise. But the spec itself contains assumptions — and some of them are wrong. Ralph Loop is a multi-round review protocol where an independent AI subagent audits every deliverable. Its exit condition borrows from mathematical convergence: one clean round proves nothing. Two consecutive clean rounds prove convergence.https://blog.chuanxilu.net/en/posts/2026/04/prd-to-tech-spec-ai-design-guardrails/Wed, 29 Apr 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/prd-to-tech-spec-ai-design-guardrails/The second Aristotle refactor had clear requirements, clean code structure, passing tests. But the async background mechanism still did not work. The problem was not in the PRD—it was in the tech spec. This article covers what a PRD should contain, what a tech spec should add, and why both are non-negotiable when AI writes your code.https://blog.chuanxilu.net/en/posts/2026/04/why-aristotle-vibe-development-needs-gear-protocol/Sat, 25 Apr 2026 00:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/why-aristotle-vibe-development-needs-gear-protocol/Aristotle v1 had a one-line requirement. The reflection task ran inside the main session, polluting 371 lines of context. This article starts from that failure and walks through why requirement gaps get amplified into systematic bias in AI-assisted development, and how structured methods close those gaps.https://blog.chuanxilu.net/en/posts/2026/04/test-doc-before-test-code-reverse-anchoring/Thu, 23 Apr 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/test-doc-before-test-code-reverse-anchoring/In AI-assisted development, tests are not just verification. They are the most precise requirement language you can give an AI. Drawing from my own failures, this article walks through the full chain from test scenario identification to test development documents, and explains why this method matters far more when the coder is an AI.https://blog.chuanxilu.net/en/posts/2026/04/managing-context-length-in-ai-coding-sessions/Sat, 18 Apr 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/managing-context-length-in-ai-coding-sessions/Someone in a group chat complained that GPT-5.4 performed worse than Doubao, ByteDance's chatbot—the model would give irrelevant answers without even reading the question. After asking some follow-up questions, I learned they had fed it many documents and the conversation had gone on for a long time. This probably wasn't the model's problem—it was context rot. The conversation had gotten so long that the model could no longer 'see' the current task clearly. This raises an overlooked problem: in the process of vibe coding or writing, how do you manage context effectively to avoid token and time wasted on model performance degradation?https://blog.chuanxilu.net/en/posts/2026/04/seven-human-ai-collaboration-patterns-in-aristotle/Thu, 16 Apr 2026 21:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/seven-human-ai-collaboration-patterns-in-aristotle/Looking back at the Aristotle project—from initial design to the GEAR protocol—I identified seven distinct collaboration patterns between myself and AI. As AI gets more capable, human judgment doesn't become less important. It becomes more critical.https://blog.chuanxilu.net/en/posts/2026/04/from-markdown-to-mcp-server-gear-protocol/Thu, 16 Apr 2026 19:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/from-markdown-to-mcp-server-gear-protocol/Aristotle's reflection rules started as a flat Markdown file — append-only, forgotten, no rollback. When dozens of rules accumulated, I realized the file wasn't enough. This started a design iteration path from append-only to Git-backed MCP Server. That path led to something called GEAR.https://blog.chuanxilu.net/en/posts/2026/04/from-scars-to-armor-harness-engineering-practice/Sat, 11 Apr 2026 01:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/from-scars-to-armor-harness-engineering-practice/The first version of Aristotle looked smooth. In practice, it exposed four architectural problems. Fixing them validated the trust model and harness engineering framework from Part 3 — every constraint encodes a trust judgment.https://blog.chuanxilu.net/en/posts/2026/04/a-trust-boundary-design-experiment/Mon, 06 Apr 2026 18:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/a-trust-boundary-design-experiment/The same reflection mechanism on different platforms, their complexity differing by an order of magnitude — but the complexity itself reveals a deeper question: when should we trust AI's judgment, and when should we step in?https://blog.chuanxilu.net/en/posts/2026/04/claude-code-reflect-different-soil/Mon, 06 Apr 2026 14:56:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/claude-code-reflect-different-soil/The same reflection mechanism lands on different platform foundations with very different landing postures and paths—from plugin installation to permission pitfalls to API concurrency, documenting the actual development process on Claude Code.https://blog.chuanxilu.net/en/posts/2026/04/aristotle-ai-reflection/Mon, 06 Apr 2026 10:00:00 +0800https://blog.chuanxilu.net/en/posts/2026/04/aristotle-ai-reflection/Installing reflection capability into AI coding assistants—when the model makes a mistake, immediately trigger root cause analysis and transform the correction into persistent rules.https://blog.chuanxilu.net/en/about/Sun, 05 Apr 2026 00:00:00 +0000https://blog.chuanxilu.net/en/about/About Chuanxilu for Skilled Homo sapienshttps://blog.chuanxilu.net/en/posts/2026/04/hello-world-the-inaugural-post/Sun, 05 Apr 2026 00:00:00 +0000https://blog.chuanxilu.net/en/posts/2026/04/hello-world-the-inaugural-post/The inaugural post of Chuanxilu of an AI Craftsman.