BUILD NOTES
问题发生以后After the problem appears
从最近的项目读回早期练习。记录项目怎样卡住，我与 AI、团队怎样处理，以及哪些问题仍然没有解决。From recent projects back to early experiments: where work stalled, how I worked with AI and teammates, and what remains unresolved.
按本篇记录的实践起点倒序，跨期项目标出时段；文章整理日期另列。Ordered by the start of the work each article covers. Periods and writing dates are shown separately.
- 01↗个人项目 · AI 协作Personal project · AI-assisted
从六面世界到最后一块地：一个 3D 游戏是怎么被“删”出来的？From a cube world to The Last Plot: a 3D game made by deleting
把六面棋局、立体贪吃蛇、生活技能教学等一堆想法删到只剩一个 5×5 的空间经济谜题，再把稀缺性和取舍做成可以亲手操作的规则。Cutting a pile of ideas — six-sided boards, 3D snake, life-skill lessons — down to one 5×5 spatial-economic puzzle, turning scarcity and tradeoffs into rules a player can operate.
- 02↗个人项目 · AI 协作Personal project · AI-assisted
个人网站：网页改好了，下载的简历为什么还是另一版？Portfolio: keeping the web resume and its download in step
从中英文网页、打印样式到单页 PDF，记录一次简历更新怎样追到实际下载文件，以及主题切换时的文字对比度问题。Following a resume revision through bilingual HTML, print styles, the downloadable PDF, and a theme-change contrast regression.
- 03↗个人项目 · AI 协作Personal project · AI-assisted
Godot 实验室：试玩能动，为什么还不能算学会了？Godot Lab: a working playground still needs a learning boundary
把一份独立 HTML 接进网站时，修复脚本转换、方向键和进度保存问题，并把浏览器模拟与真实 Godot 实操分开。Integrating an HTML learning lab exposed script conversion, keyboard, and storage failures, alongside a gap between simulation and editor practice.
- 04↗个人项目 · AI 协作Personal project · AI-assisted
二维码实验室：自己能解开，为什么别的扫描器读不出？QR Studio: when the encoder and decoder agree on the same mistake
一次二维码页面接入暴露了掩模运算和版本信息的位置错误。我们用独立解码器与参考矩阵，补上同实现往返测试的盲区。Mask arithmetic and version-information placement escaped internal round trips, prompting independent decoders and reference-matrix checks.
- 05↗个人项目 · AI 协作Personal project · AI-assisted
声音手帖：重建练习之前，先保护已经学过的记录Sound Notebook: rebuilding practice without inventing learning history
从旧版自标记迁移到独立假名卡与 FSRS，再加上笔顺示范。复盘旧记录、音频失败、开源许可和动画生命周期怎样分别处理。Migrating self-marked progress to independent kana cards and FSRS, then adding stroke playback while protecting history and handling audio failures.
- 06↗个人项目 · AI 协作Personal project · AI-assisted
会务实践：一处资料变更，怎样跟到最后一个交付物Conference practice: following one revision through to the last deliverable
从串场 PPT、人像和配音制作，到资料变更与会后交付，记录真实任务怎样逐渐沉淀成可复用的方法，以及哪些检查仍要由人完成。How real assignments involving slides, portraits, narration, and delivery became reusable methods, with clear limits on what AI and file checks can confirm.
- 07↗个人项目 · AI 协作Personal project · AI-assisted
StudyMate：诊断做完以后，下一步到底能做什么？StudyMate: turning diagnostic feedback into a task someone can start
从只有分科时间的学习计划，推进到第一种可恢复的写作训练；记录评分截断、失败重试和辅助改稿与诊断成绩的边界。Moving beyond time allocations to a recoverable writing task, while handling incomplete model feedback and keeping assisted revision separate from diagnosis.
- 08↗个人项目 · AI 协作Personal project · AI-assisted
Planbox：能找到原文，不等于信息已经可信Planbox: finding the source does not make information trustworthy
活动整理原型从抽取与人工确认，推进到时间精度、更正历史和撤回；新的采集缺陷也提醒我，来源归属必须先于自动排期。An event-information prototype adds date precision, correction history, and retraction, while new intake defects show why source attribution must come before automatic scheduling.
- 09↗个人项目 · AI 协作Personal project · AI-assisted
CyberCare：把 AI 识别结果留在草稿里CyberCare: keeping AI extraction as a draft
从海报缺失的报名方式，到语音中的地点误识别，记录一次信息可及性原型如何保留未知、让人核对，并区分技术跑通与真正有用。From missing registration details to a misheard location, an information-accessibility prototype explores explicit unknowns, human review, and the gap between a working system and useful help.
- 10↗个人项目 · AI 协作Personal project · AI-assisted
ArrivalReady：从 AI 预审到人工确认与整改复测ArrivalReady: From AI pre-check to human confirmation and retest
从收紧比赛范围，到修复报告分数消失、整改入口为空和复测对比不显示，记录我怎样与 AI 协作完成一次完整流程验收。A constrained prototype review of evidence, AI pre-checks, human confirmation, deterministic scoring, remediation, and retesting—with each claim kept within its evidence.
- 11↗个人项目 · AI 协作Personal project · AI-assisted
OneCase：接口不报错了，却多出一件不该办的事OneCase: The API stopped failing, but created a task that should not exist
从消息拆分和事务确认，到模型跟随编造指令、把问候语变成草稿，记录社区事项原型暴露的问题与尚未解决的取舍。A community-casework prototype about message splitting, transactional confirmation, model instruction-following failures, and the tradeoffs that remain unresolved.
- 12↗个人项目 · AI 协作Personal project · AI-assisted
H2 Manual Copilot：引用点得开，答案就已经可靠吗？H2 Manual Copilot: a clickable citation still needs a trustworthy answer
在设备手册问答原型中，把页面、检索、引用与模型质量拆开验收；记录替身模型和合成手册能证明什么，以及回滚暴露的枚举残留。Separating retrieval, source navigation, model quality, and rollback checks in an equipment-manual prototype built with synthetic documents and a model substitute.
- 13↗个人项目 · 多 Agent 实验Personal project · multi-agent experiment
TeachOps：四个 Agent 跑完，审计仍然给了 FAILTeachOps: four agents completed the run, and the audit still failed
一次真实多 Agent 运行发现修订稿漏掉关键学情误解，导师据此附条件批准；同时检查缺证据时下游是否真的停下。A live multi-agent run missed a key learner misconception, prompting conditional human approval and a separate missing-evidence stop check.
- 14↗个人项目 · AI 协作Personal project · AI-assisted
EasyPhone：教程写出来了，老人从哪一步开始？EasyPhone: The tutorial was written—but where should an older user begin?
把教老人用手机的宽泛目标缩到一步指导，再用确认、风险停止和求助卡处理原型中的边界。A constrained phone-help prototype: one actionable step, confirmation, risk stops, clarification, and a family help card, with no claim of older-adult validation.
- 15↗个人项目 · AI 协作Personal project · AI-assisted
LOOM 的几次演进：画布能连起来，执行链为什么还会断？LOOM iterations: connecting a canvas is only the start of execution
从 LOOM、AgentFlow 到 LOOM-PLUS，沿 PDF 解析接口变更与向量维度不匹配，回看整合已有代码时需要重新核查的契约。Following PDF-parser API changes and embedding-dimension mismatches through the integration of LOOM and AgentFlow into LOOM-PLUS.
- 16↗个人项目 · AI 协作Personal project · AI-assisted
WechatArticle：两次内容生产与跨平台发布的复盘WechatArticle: Two content runs and a cross-platform publishing workflow
从补上即刻交接提示词，到统一小红书配图模板、清理生成文件，记录这个自用写作工具怎样逐步补齐发布前的工作。A review of how a personal writing tool added a Jike handoff prompt, shared Xiaohongshu visual templates, and separated generated files from retained content.
- 17↗个人项目 · AI 协作Personal project · AI-assisted
JobMatch：岗位表合并了，报名人数却可能跟错岗位JobMatch: a successful spreadsheet import can still join the wrong records
把选岗工具收窄到浏览器内处理以后，难点落在表头与关联键；源码复核又发现前导零和重复键并未像旧文档说得那样解决。A browser-only job filter moved complexity into headers and join keys; source review also exposed gaps behind earlier normalization claims.
- 18↗团队项目 · 模块协作Team project · module collaboration
EventAccount：我负责发现页，为什么审查追到了报名事务？EventAccount: how reviewing discovery reached registration transactions
团队按模块交付前后端。我在个人中心、发现页和落地页的工作中，遇到公开数据边界、请求竞态与候补递补的一致性问题。A team project with end-to-end module ownership exposed public-data boundaries, stale discovery results, and waitlist consistency across modules.
- 19↗个人项目 · AI 协作Personal project · AI-assisted
角色聊天：发了一张图，AI 怎么把人设忘了？Role chat: keeping a character across an image message
早期多模态聊天把看图与角色回复拆成两个阶段，也暴露了额外延迟与中间描述丢信息的代价。An early multimodal chat separates image description from character response, introducing a clearer handoff and additional failure points.
- 20↗个人项目 · AI 协作Personal project · AI-assisted
最早的备忘录：读不出的数据，为什么不能当成空列表？The early memo app: unreadable data must not become an empty notebook
从 2025 年的 Flask 小应用回看备份恢复、置顶排序与测试隔离：一次看似友好的异常兜底，可能把恢复机会也覆盖掉。Revisiting a Flask app started in 2025 through recovery, pinned ordering, and test isolation, including a dangerous empty-list fallback.