https://yistar.jiezeng2004.me/en Working across genetics, developmental biology and bioinformatics, and turning recurring AI workflow problems into local-first, inspectable tools. en https://yistar.jiezeng2004.me/en/build-log/vibe-coding-real-cost https://yistar.jiezeng2004.me/en/build-log/vibe-coding-real-cost Sun, 01 Feb 2026 00:00:00 GMT The concept of Vibe Coding is appealing: go with the flow and let AI write code for you. But in real projects, I've found that "going with the flow" is not actually low-cost. Real costs include: Verification costAI-generated code requires human review. Not all code reveals problems at a glance — some bugs require running tests or even manual testing to discover.Context costAI doesn't understand your full project context. It might use wrong library versions, ignore existing conventions, or reimplement existing functionality.Boundary costAI might modify files it shouldn't or introduce unnecessary dependencies. Without clear workspace scope limits, risks multiply.Audit costWhen AI output has problems, you need to trace back: who generated this code? Based on what context? Which files were modified? My solution is to establish a verification workflow: After each AI code generation, review the diff firstRun existing tests to confirm no existing functionality is brokenManual testing on critical pathsRetain task trail evidence The "vibe" in Vibe Coding should not replace rigorous engineering practices. AI makes coding faster, but the verification stage cannot be skipped. This is exactly the core philosophy of PatchWarden: making AI agent output auditable, verifiable, and reversible. https://yistar.jiezeng2004.me/en/build-log/chatgpt-local-agent-collaboration https://yistar.jiezeng2004.me/en/build-log/chatgpt-local-agent-collaboration Thu, 01 Jan 2026 00:00:00 GMT After months of practice, I've gradually formed a collaboration pattern between ChatGPT and local agents. What ChatGPT excels at: Requirement discussion and solution designComparing and weighing multiple approachesConcept explanation and knowledge queriesCode review and logical reasoning What local agents excel at: Executing modifications in the actual codebaseBatch file operationsRunning tests and verificationGenerating structured diff reports My workflow: Clarify requirements in ChatGPT, discuss multiple approachesAfter determining the final approach, describe the task in structured formatHand the task to a local agent for executionReview the agent's diff and verification reportsDecide whether to accept the changes Throughout this process, the human remains the decision-maker. AI is a tool, not a replacement. Failure experiences are equally important. Some tasks I overestimated the agent's understanding, leading to results deviating from expectations. Some tasks I didn't define boundaries clearly enough, causing the agent to modify files it shouldn't have. These failures made me realize: the core of AI agent collaboration isn't making AI do more, but making humans think more clearly. https://yistar.jiezeng2004.me/en/build-log/why-patchwarden https://yistar.jiezeng2004.me/en/build-log/why-patchwarden Mon, 01 Dec 2025 00:00:00 GMT When using ChatGPT to discuss code approaches daily, I encountered a recurring problem: the discussion phase and execution phase are disconnected. ChatGPT provides excellent solution suggestions, but it cannot operate on my local codebase. Meanwhile, local agents like Codex and Claude Code have strong execution capabilities but lack the context from the discussion phase. When I copy solutions to local agents for execution, I often encounter: Unclear workspace scope — agents might modify files they shouldn'tLack of structured verification reports for execution resultsNo traceable audit trail for the entire process PatchWarden was created to solve this. It establishes a controlled channel between ChatGPT and local agents: extract tasks, limit scope, execute, return diffs and verification reports. This is not a general-purpose remote shell, but a task execution framework with clear boundaries. Every step is recorded, every change is traceable. The process of building this tool is itself a practical exploration of "how humans and AI agents should collaborate."