https://www.dengmengmian.com/en/writingExploring how individuals build software, work with AI agents, and design durable systems for work and life.enMafan Lab / ForgeKit© 2026 MafanThu, 27 Aug 2026 14:23:08 GMThttps://www.dengmengmian.com/en/writing/proving-agent-improvement-beyond-traditional-testshttps://www.dengmengmian.com/en/writing/proving-agent-improvement-beyond-traditional-testsFri, 21 Aug 2026 05:59:06 GMTMafanTraditional tests only prove the code isn't broken; they can't evaluate whether the Agent has truly improved. Eval is needed: run a fixed task set with unified scoring rules repeatedly, comparing success rate, cost, time, and regressions. The two systems complement each other; before release, pass traditional tests first, then run Eval, to avoid optimizing by feel.Agent开发AI评测Eval大模型应用开发https://www.dengmengmian.com/en/writing/building-coding-agent-just-beginninghttps://www.dengmengmian.com/en/writing/building-coding-agent-just-beginningFri, 21 Aug 2026 05:56:25 GMTMafanBuilding a coding agent taught me that calling the model is just the beginning. In long tasks, state management, tool logging, observability, and failure recovery are key. Chat is just the entry point; the system must act like a runtime to keep progressing and validating results. Reliability often matters more than a one-off impressive answer.效率提升产品思维AI编程https://www.dengmengmian.com/en/writing/chatgpt-codex-merger-terms-updatehttps://www.dengmengmian.com/en/writing/chatgpt-codex-merger-terms-updateFri, 21 Aug 2026 05:54:15 GMTMafanAfter ChatGPT and Codex merged, the protocol changed again: tool definitions moved into additional_tools, and third-party models that don't receive tools may fake calls. The author's local gateway AgentGate has adapted, restoring tools and forwarding upstream; the merged desktop client can still point to localhost to call DeepSeek and other models.本地网关协议转换AI编程https://www.dengmengmian.com/en/writing/my-self-made-toolhttps://www.dengmengmian.com/en/writing/my-self-made-toolFri, 21 Aug 2026 05:51:22 GMTMafanReviewGate is an open-source MIT-licensed pre-merge filtering tool that screens for high-risk issues from perspectives like security, performance, logic, and AI Smell, categorizing them as block or warn. It prefers to fail rather than give a false green light. It supports --fix, CI integration, and false-positive sharing, and is read-only by default. It's recommended to observe first, then gradually enable.需求验证AI编程https://www.dengmengmian.com/en/writing/midlife-fitness-solitude-thinking-timehttps://www.dengmengmian.com/en/writing/midlife-fitness-solitude-thinking-timeFri, 21 Aug 2026 05:47:43 GMTMafanAfter middle age, the meaning of fitness goes beyond physical exercise; it becomes a deliberate way to carve out uninterrupted personal time. It allows one to temporarily step away from work, family, and fragmented information, restoring attention and order, and reclaiming quiet space for independent thought. This ability to 'return to oneself' is more important than mere physical management.一人做事节奏感https://www.dengmengmian.com/en/writing/overestimating-execution-underestimating-direction-and-rhythmhttps://www.dengmengmian.com/en/writing/overestimating-execution-underestimating-direction-and-rhythmTue, 28 Jul 2026 16:25:31 GMTMafanWhen working alone, it's easy to overestimate execution and underestimate direction and pace. Execution is visible, but if the direction is wrong or the pace is chaotic, the harder you work, the further you may drift from key results. What truly matters is constantly judging the main line, stage priorities, and what's most important, avoiding being busy but not making real progress. A sense of direction and pace matter more than sheer effort for steady long-term progress.执行力方向感节奏感https://www.dengmengmian.com/en/writing/i-like-codex-but-it-cant-use-deepseek-so-i-made-a-local-gatewayhttps://www.dengmengmian.com/en/writing/i-like-codex-but-it-cant-use-deepseek-so-i-made-a-local-gatewayTue, 28 Jul 2026 16:25:31 GMTMafanThe author likes Codex but wants to use DeepSeek, and finds the protocols incompatible. So they developed a local gateway called AgentGate. This tool converts Codex's Responses API requests into DeepSeek's Chat Completions format, and further implements intelligent routing: text-only tasks automatically go to DeepSeek, while multimodal tasks with images switch to KimiCoding. This solves the fragmentation of tool and model protocols in AI coding, letting users focus on development.本地网关协议转换AI编程https://www.dengmengmian.com/en/writing/why-projects-fail-not-due-to-slow-development-but-scope-creephttps://www.dengmengmian.com/en/writing/why-projects-fail-not-due-to-slow-development-but-scope-creepTue, 28 Jul 2026 16:25:31 GMTMafanProject unprofitability is often attributed to slow development, but the more critical factor is scope creep. Slow development affects efficiency, while scope creep undermines the profit structure. Small requirements accumulate, causing scope creep, distorted quotes, and gradually eroded profits. Holding the line on boundaries is more important than improving efficiency: clarify what's not included, designate decision-makers, watch out for accumulating small changes, and define clear acceptance criteria.边界失控项目利润需求蔓延https://www.dengmengmian.com/en/writing/how-to-decide-if-a-task-is-worth-delegating-to-aihttps://www.dengmengmian.com/en/writing/how-to-decide-if-a-task-is-worth-delegating-to-aiTue, 28 Jul 2026 16:25:30 GMTMafanDeciding whether to delegate a work task to AI hinges on distinguishing between "can delegate" and "should delegate." The author proposes five criteria: tasks with high repetition but low judgment density are suitable for AI; tasks that would weaken your understanding of the problem require caution; tasks with high error costs should not be delegated; tasks that allow quick verification and correction are suitable; and tasks that require a stance rather than mere output should be done yourself. The core is AI工作边界效率与失控判断标准https://www.dengmengmian.com/en/writing/5-things-to-evaluate-a-requirementhttps://www.dengmengmian.com/en/writing/5-things-to-evaluate-a-requirementTue, 28 Jul 2026 16:25:30 GMTMafanThis article shares five core principles for judging whether a feature is worth building: confirming the problem truly exists rather than being imagined, assessing whether the pain point is strong enough, distinguishing whether the feature drives results or merely fills a sense of completeness, determining whether the timing is right, and whether the feature can bring verifiable changes after completion. It emphasizes that product success lies not in feature accumulation but in precise filtering and priorit需求判断产品优先级痛点分析https://www.dengmengmian.com/en/writing/ai-era-hardest-part-is-not-development-but-persisting-in-doing-the-right-thingshttps://www.dengmengmian.com/en/writing/ai-era-hardest-part-is-not-development-but-persisting-in-doing-the-right-thingsTue, 28 Jul 2026 16:25:30 GMTMafanAI lowers the barrier for individual development, but the real challenge for solo workers is continuously doing the right things: choosing direction, managing pace, making trade-offs, and maintaining persistence. AI makes 'being able to do' easier, but makes 'what to do' more critical, easily leading to scattered focus on things that seem right but aren't key.一人做事方向选择AI门槛https://www.dengmengmian.com/en/writing/confirming-requirements-boundaries-before-technical-solutionshttps://www.dengmengmian.com/en/writing/confirming-requirements-boundaries-before-technical-solutionsTue, 28 Jul 2026 16:25:30 GMTMafanWhen taking on a project, confirm the requirement boundaries before the technical solution. The boundaries determine what the project does, what it doesn't do, who has the final say, how changes are handled, and what delivery is based on. Unclear boundaries can cause the project to be dragged down by scattered requirements, losing control of profit and pace. Confirming the core goal, defining the scope, specifying the decision-maker, agreeing on change rules, and delivery standards are prerequisites for pro需求边界项目失控变更管理https://www.dengmengmian.com/en/writing/ai-rewriting-workflow-not-every-step-worth-ithttps://www.dengmengmian.com/en/writing/ai-rewriting-workflow-not-every-step-worth-itTue, 28 Jul 2026 16:25:30 GMTMafanAI is changing my workflow, but not every step is suitable for it. I've learned to first judge which tasks are worth AI's involvement, such as drafts, information organization, and repetitive work, while key decisions, final structure, and business understanding still require my personal control. This division of labor improves efficiency but avoids blind reliance leading to loss of control.AI工作流效率提升人机协作https://www.dengmengmian.com/en/writing/cutting-illusions-not-features-in-product-developmenthttps://www.dengmengmian.com/en/writing/cutting-illusions-not-features-in-product-developmentTue, 28 Jul 2026 16:25:30 GMTMafanIn the past, when building products, I always preferred to add features. Later, I gradually realized that many products don't die because they can't be built, but because they think too much. In the early stages of a product, what should be cut first is often not pages or flows, but those unverified illusions that are already taken for granted.产品思维需求验证创业反思