Tag
作者指出 AI 功能在 demo 阶段表现优异却在生产环境中失灵的常见原因：示例是由人工挑选的理想输入。建议在写第一个 prompt 前先收集 50-100 个真实（尤其是糟糕的）输入并定义理想答案，用作回归评测集。
Obie Fernández, CTO of ZAR, introduces 'Impatient Programming' (IMP), a revisiting of Extreme Programming's 29 rules for teams building software with coding agents, along with guardrail tools and a directory of 539 related tools.
文章探讨了 AI agent 系统中「任务完成」信号不可靠的问题，并提出三层验证框架——回执、读回校验和独立检查——以证据驱动的方式定义何为真正的完成。
推文引用了Elon Musk关于“超智能的超强说服力会让人一击即中”的评论，并指出许多工程师仅仅因为AI模型的解释听起来合理就直接合并代码修复。
Anthropic 分享 Claude Code 团队如何在 AI 写了约 90% 代码的情况下进行代码评审：通过多层自动化检查、把人工注意力集中在高风险环节，并在 PR 数量翻倍的同时将回滚率降低 40% 以上。
OpenAI's official blog revealed how its core engineering team uses the GPT-6 series (including GPT-6 Astra) and Codex for large-scale codebase maintenance, merging over 1,400 cleanup PRs in two weeks, deleting 800k+ lines of code, and sharing the actual production prompts used for dead-code cleanup, feature-flag pruning, CI optimization, flaky test fixes, and agent instruction self-inspection.
Gergely Orosz asks engineering leaders to share their team's PR output before and after adopting AI tools, along with per-engineer AI spend, to compile and share anonymized industry benchmarks on AI's impact on developer productivity.
Kent C. Dodds shares 15 product development principles behind how he ships features, published via MEGA.dev's "How Kent Ships Features Without Reading the Code" piece.
A practitioner argues that the real difficulty of building AI agents is not the model but data quality, failure-path design, process discovery, scope discipline, and handoff — sharing concrete checks and rules for deciding what to automate.
This article revisits a blog post by the Manus AI team from a year ago on context engineering practices, summarizing six key lessons for building AI Agents, such as context management, memory recovery, and error handling.
Tobi Lütke discusses his approach to simplifying internal tools at Shopify by making quick architectural decisions, emphasizing founder-led interventions to avoid complexity and technical debt.
This article emphasizes the need for trust in AI-generated code within engineering teams and outlines practices such as accountability, coding guidelines, and deterministic tooling to uphold code quality.
Anthropic has released a 40-minute comprehensive guide on Loop Engineering, covering topics from building basic loops to advanced goal-setting in AI agents, aimed at changing how developers approach loop construction.
Dropbox discusses strategies to improve infrastructure efficiency in response to growing AI demand, focusing on system-level optimization rather than just adding more hardware.
This describes a set of agent skills for real engineering with coding agents like Claude Code and Codex, aimed at fixing common issues in AI-assisted development.
Addy Osmani discusses practical loop engineering for AI agents, covering goals, autonomous feedback cycles, and the use of tools like Claude Code and Codex for managing parallel agent tasks.
Kent C. Dodds highlights Sunil Pai's explanation of product engineering's importance, and Pai announces his new role as builder in residence at Cloudflare to focus on identifying emerging trends.
Simon Willison highlights Sophie Alpert's essay arguing there are no lossless transformations of natural-language text; AI-assisted rewrites can lose the author's intended meaning, so engineers must stand behind every sentence they publish.
AGENTS.md summarized by Vercel engineers can save you 80% of tokens in project development. The core is to reduce AI's ineffective iterations and token consumption through engineering decisions and solution design.
An analysis of common reasons why software factories fail, arguing that focusing solely on harness engineering is insufficient for success.