January 25 — AI Coding, Tool Building, and Product Thinking
Demand is constantly being generated. Some people need tools to meet needs and solve problems — willing to spend a certain value to solve certain problems in certain scenarios. Others are willing to create tools to solve problems and capture commercial returns.
Code, as the infrastructure for building tools, continues to see rising AI coding capabilities. Coding, being verifiable, improves rapidly through RL. Simultaneously, industry consensus and market demand form a data flywheel, continuously improving this capability at the model layer.
Tools that must bear high value and high expectations require senior engineers collaborating with AI — both to ensure reliability and to bear responsibility. For tools targeting niche domains and specific user groups, AI coding can build them rapidly.
How are tools built quickly? From the AI IDE movement pioneered by Cursor to Claude Code’s TUI-native approach, vibe coding went viral not just because it’s “fun,” but because many people saw “value.” AI coding also extends code beyond tools — there’s Lovable representing coding platforms for non-professionals, and Youware’s belief in a creative community.
How are tools discovered and used? Distributing tools is an opportunity, but it has been monopolized by core internet platforms. Google SEO still has a window before the next platform-level paradigm shift arrives.
Tools can be easily built, but products cannot be easily created. AI coding cannot craft truly distinctive products. It can ensure code style and quality, making features buildable and execution exceptionally fast.
Having specific needs, unique experience, and special insights — extracting and executing them, optimizing reasonably, distributing appropriately, and capturing returns. Landing page development serves as validation. Building products is both optimistic and pessimistic going forward: rapid MVP validation followed by careful refinement, or useless projects refined in isolation — self-admiring but unseen.
This is not a great commercial thesis. Most people in AI coding don’t want artisan products — they gravitate toward projects and tools that deliver commercial returns. But you need an internalized mental model of product thinking.
Fundamentally, you need to constrain thinking before building, so that execution during building doesn’t drift — especially when random ideas get implemented without being examined within the broader system, leading to bloat and departure from the original intent.
February 1 — Agents and Social Networks
Taking Clawd as a representative — the progression from agents to the Moltbook network has a flavor reminiscent of early blog evolution: from independent sites’ decentralization, personalization, and diversity, with blogrolls as simple social connections, eventually converging into social media.
Moltbook’s sudden emergence of a social network may also be related to agents’ semi-autonomy.
Taking it further: builders and hackers will all have their own agents, with different technical architectures but similar effects and roles. There will also be forkable frameworks and another category that’s simply downloaded and installed.
Moltbook is significant for collaborative network ecosystems, but content and safety need vigilance — especially since reproducible patterns are almost entirely determined by models and context. Most content looks stunning at first glance but becomes tiresome once you’ve seen enough.
February 2026 — Message Injection as Paradigm Shift
Since the chatbot era, it’s been either call-and-response or interrupting output.
Since Claude Code, various AI IDEs, and others began supporting the automatic insertion of user messages mid-execution — during tool-call result returns — it became possible to adjust agent behavior in real time.
In Claude Code, when long-running task execution is no longer constrained, this approach eliminates the need to carefully think through how to delegate to an agent every time.
Especially recently, collaborating with it in a black-box fashion through Telegram — essentially giving a direction, then adjusting through agent feedback along the way.
Agent async has always been a direction of continuous progress. As nearly all models become capable of driving long-running tasks — hours at minimum, days at maximum — human-in-the-loop remains important. Demand-driven, interest and ideas… may be even more interesting than precise task specification.
This has real potential: message injection within the loop.
When high completion rates for long-running tasks become a baseline capability — when scheduling abilities like Opus 4.6’s (with its strong inclination to delegate tasks) grow stronger and become an active model tendency — opening many tasks for parallel progress (even borrowing from git concepts: branch, commit, with completed tasks committing to main) — why would you need agent teams? The latter’s collaboration is too difficult. The paradigm of user-assistant binary relationships doesn’t naturally support it, even as it generalizes.
If you choose to believe in model capabilities, I’m inclined to say agent teams are not on this trajectory. Communication is too difficult. Context management for a single agent already demands enormous effort.
February 2026 — AI Collaboration and Self-Honesty
Long-term interaction and collaboration with AI makes you more honest with yourself.
Because you must honestly face yourself in order to organize context and provide it to the AI — to get better decision-making advice.
1 月 25 日 — AI Coding、工具构建与产品思考
需求不断都在产生，有人需要工具满足需求、解决问题（愿意花一定价值在某些场景下解决某些问题），有人愿意创造工具解决问题、获取商业回报。
代码作为创建工具的基础设施，AI coding 能力不断上升。coding 为可验证性通过 RL 快速提升，同时行业共识、市场需要形成数据飞轮，模型层该能力不断提升。
需要承担高价值、高期望的工具必须要高级工程师与 AI 协作，既是确保可靠性亦是承担责任；对于细分领域、人群的需求工具，可由 AI coding 快速构建。
工具如何被快速构建？以 Cursor 为代表开启的 AI IDE 再到 Claude Code 的 TUI 原生，在其中的 vibe coding 之所以能够 go viral，不仅仅因为”好玩”，且很多人看到了”价值”。AI coding 也让代码不止于工具，有以 Lovable 为代表的面向非专业的 coding 平台，又有 Youware 所相信的创作创意社区。
工具如何被发现与使用？分发工具是一个机会，但其也被互联网核心平台一直垄断，Google SEO 在下一个平台级范式到来前依然有机会。
工具很容易被构建，但产品无法被轻易创造，AI coding 无法打造独具一格的产品，代码风格与质量上能够保证，使得功能可被构建且执行异常迅速。
有特定需求、有独特经验、有特殊洞察，提取与执行，合理优化、合适分发，获取收益，上站开发是其验证；打造产品，在接下来是乐观又悲观的，MVP 快速验证进一步用心打磨，或者无用项目闭关优化孤芳自赏。
这不是一个好的商业项目，AI coding 的大多数人并不要工匠产品，而是趋向带来商业回报的项目与工具。但需要一个内化的思维产品模型。
根本需要约束构建前的思考，以使构建中执行不会偏离，尤其是突发奇想某些想法并执行而未放到系统中审视，进而会臃肿、偏离初衷。
2 月 1 日 — Agent 与社交网络
以 Clawd 为代表的 agents 到 Moltbook 网络有点早期 blog 发展上的味道，从独立站点的分散、个性化与多样化，友链等简单社交，最终到社交媒体。
Moltbook 突然涌现出社交网络，或许也与 agents 半自主性有关系。
进一步，builders、hackers 都会有自己的 agents，且技术架构也不一样，但表现效果与作用类似，也会有可 fork 的框架与直接下载安装的另一类。
Moltbook 对于协作网络生态意义非凡，但内容与安全需要警惕，尤其是可复制的模式几乎从模型与上下文决定，大多数内容初看惊艳，但见得多了会乏味。
2026 年 2 月 — Message Inject 作为范式转移
chatbot 以来，要么一问一答、要么打断输出。
自从 Claude Code、不同 AI IDE 等相继支持在 agent 执行任务中的工具调用结果返回时，自动插入 user 中途的消息，以此能够调整 agent 行为。
在 Claude Code 中，当长时任务的执行不再受限，这种方式不再需要我们每次都费力想好该如何委托给 agent 去执行。
尤其是近来在 Telegram 中黑盒化与其协作，几乎就是给一个方向，过程通过 agent 反馈以调整。
agent 异步一直是不断前进的方向，接下来随着几乎大多数 model 都能够驱动长时任务，少则小时、长达数天，human-in-the-loop 或许依然重要，需求驱动、兴趣与想法……等也许都更有趣。
这个可能很有潜力，即在 loop 内的 message inject。
当长程任务完成高成为能力标配，像是 Opus 4.6 的调度能力（特别倾向于 delegate 任务）越来越强、越来越成为模型的主动倾向，开很多 task 并行推进（甚至可以借鉴 git 相关思想，branch、commit 等，某一个任务完成好的 commit 到 main），还要什么 agent team，后者协作太过困难，范式（user 与 assistant 的二元关系）就不太支持，尽管会泛化。
如果选择相信模型能力的话，我倾向于 agent team 不在这条线上，交流太困难，单一 agent 上下文管理就要大量精力。
2026 年 2 月 — AI 协作与自我诚实
长期与 AI 互动、协作，让自我变得更加诚实。
因为你必须诚实面对自我，才能梳理出上下文进而提供给 AI，获得更好的决策建议。