AI Workflow · Ops Tooling · Human-in-the-loop · Workflow Design

Semi-automated Job Decision Workflow

AI 求职决策助手 · 半自动化 JD 匹配工作流

2026

01 | 项目概述 Project Summary

一个围绕「建筑设计师转 AI 行业求职」真实痛点搭建的半自动化决策工作流。通过 Claude API 实时解析岗位 JD，匹配三版差异化简历与招呼语版本，输出结构化「投递决策卡」。

A semi-automated decision workflow built around a real pain point — transitioning from architectural design into the AI industry. It uses the Claude API to parse job descriptions in real time, match against three differentiated resume and outreach versions, and output a structured "application decision card."

工作流核心定位：不替代人做判断，而是把"识别同名异向、版本匹配、风险归类"这类结构化但重复的判断成本降到趋近于零，让使用者把精力留在真正需要权衡的取舍上。

The core positioning is not to replace human judgment, but to drive the cost of structured-yet-repetitive decisions toward zero — letting the user keep focus on the trade-offs that actually require thinking.

两天内累计判断 108 个岗位，平均决策时间从 5 分钟 / 个降至 30 秒 / 个。

Across two days, 108 positions were processed; per-decision time dropped from roughly 5 minutes to 30 seconds.

02 | 为什么做这个 Why I Built This

转行求职过程中遇到三个交叠的真问题。

Three overlapping pain points encountered during this career-transition job search.

同名异向 / Same Title, Different Reality

同样叫"AI 产品运营"的岗位，实际工作内容差异极大——有的偏投流，有的偏内容，有的偏 prompt 工程。光看 title 无法判断方向。

Jobs called "AI Product Operations" can mean ad performance, content ops, or prompt engineering. Title alone cannot tell.

重复决策成本高 / High Repetitive Decision Cost

每个岗位都要经历"读 JD → 判断方向 → 选简历版本 → 写招呼语 → 决定投不投"。一晚最多深度判断 10 个岗位，效率瓶颈在判断而非投递。

Each posting required reading JD → deciding direction → picking resume version → writing greeting → final go/no-go. Throughput was capped at the decision step, not the application step.

转行简历容易被秒拒 / Transition Resumes Get Filtered Fast

建筑专业 + AI 方向求职，HR 看到第一眼时往往直接划走。投出去无回复是常态。

Coming from architecture into AI, recruiters often dismiss the resume at first glance. Silent rejections are the norm.

第三点工具解决不了。所以这个工具的目标不是"提高回复率"，是在"看不到回报的求职阶段，依然能按结构化节奏继续投递，不被疲惫吞没"。

The third pain point is not solvable by tooling. So the goal of this workflow is not to "increase reply rate" — it is to keep moving at a structured pace during a phase where feedback is delayed, so that fatigue does not become the bottleneck.

03 | 工作流怎么设计 Workflow Design

JD 文本输入（截图 OCR / 复制粘贴）
        ↓
Claude API 实时解析（嵌入个人背景 + 判断模板）
        ↓
五维度判断
  · 岗位偏向（应用层 / support / workflow / ops vs 工程 / 算法）
  · 核心要求（文档 / 协同 / 用户支持 / 工程实现）
  · 个人匹配点
  · 风险点（工程深度 / 英文要求 / title 名实不符）
  · 综合结论（主投 / 可冲 / 暂不主攻）
        ↓
版本匹配
  · 简历 A / B / C（产品运营 / 平台支持 / 应用运营）
  · 招呼语 A / B / C（与简历同向）
        ↓
结构化决策卡输出（10 字段）
        ↓
本地双轨留存
  · Markdown：每岗一份，便于回看
  · CSV 台账：用于后续回复率统计

设计原则 / Design Principles

Workflow Design —— 把重复人工判断转成可复用的结构化输出，每张决策卡 10 个字段固定，便于横向比对。

Convert repetitive judgment into reusable structured output, with 10 fixed fields per card to allow horizontal comparison.

Human-in-the-loop —— 工具只判断不投递，"立即沟通 / 投递 / 拒绝"由使用者手动完成。这一点是有意为之，07 节展开。

Tool decides, human applies. "Initiate chat / submit / decline" remain manual actions — a deliberate boundary, expanded in section 07.

降级容错 / Graceful Degradation —— 如果 JD 提取失败或方向无法识别，决策卡会显式输出"信息不足"或"无可用简历版本"，不强行做判断。

When JD extraction fails or direction cannot be inferred, the card explicitly outputs "insufficient information" rather than forcing a decision.

04 | 技术路径选择 Path Selection

在到达当前命令行半自动版本之前，先后实践 / 评估过三条技术路径，每条因为不同原因调整方向。完整记录构成了项目的 Failure Postmortem 档案——这是 ops 工作里"决策可追溯"的核心实践。

Before arriving at the current CLI semi-automatic version, three technical paths were explored and adjusted. The full record forms the project's Failure Postmortem archive — a core practice of "traceable decision-making" in ops work.

路径 1 | BOSS CLI 接口自动化 / API Automation

尝试通过开源 boss-cli 工具搭建"搜索 → 详情 → 筛选 → 招呼"全自动链路，包裹一层 safe wrapper 严格阻止越权操作（cookie 读取 / 自动登录 / 自动投递）。

Attempted full automation via open-source boss-cli, wrapped in a safe layer that strictly blocked over-privileged actions (no cookie reads, no auto-login, no auto-apply).

结果：单次只读 search 跑通过；详情接口受 session 时效限制；批量 search 触发平台风控信号（"风控" marker）。Wrapper 按既定规则立即 STOP，避免了账号风险。

Outcome: single-shot read-only search worked; detail endpoint hit session-expiry constraints; batch search triggered the platform's risk-control marker. The wrapper executed an immediate STOP per design, protecting the account.

结论：BOSS 平台对高频自动化不友好，接口路径不适合作为主流程。但 wrapper 的 STOP 机制和 diagnostics 设计被证明是有效的安全边界——这部分思路被保留下来，体现在后续工具的"只读不投"原则里。

Conclusion: BOSS is not friendly to high-frequency automation; the API path cannot be the main pipeline. But the wrapper's STOP and diagnostics design proved an effective safety boundary — this thinking carries into the "read-only, no actions" principle in later tools.

路径 2 | 油猴脚本浏览器浮窗注入 / Browser Overlay Injection

打算在 Boss 直聘网页右侧实时注入决策卡浮窗。脚本完整生成（含 Tampermonkey + OpenRouter API + Shadow DOM 隔离 + GM_setValue 安全存储）。

Planned a real-time decision card overlay injected into Boss's web pages. Full script generated (Tampermonkey + OpenRouter API + Shadow DOM isolation + GM_setValue secure storage).

结果：发现 Boss 直聘有强反 DevTools 检测（打开开发者工具直接踢出页面），且新版"列表 + 预览面板"双栏布局让 SPA 路由检测失效，调试成本远高于预期收益。

Outcome: discovered strong anti-DevTools detection on Boss (opening DevTools immediately ejected the page) and a new dual-pane "list + preview" layout that defeated SPA route detection. Debugging cost far exceeded expected value.

结论：浮窗思路本身是对的，但选错了战场。浮窗版的代码与架构被保留作为后续 LinkedIn / 猎聘适配的基础——这两个平台反爬强度低得多。

Conclusion: the overlay idea is right; the battlefield was wrong. The overlay code and architecture are preserved as a base for future LinkedIn / Liepin adaptations, where anti-scraping is far lighter.

路径 3 | 命令行半自动 / CLI Semi-automatic（当前主流程 Current Main Pipeline）

回归最朴素方案：终端输入 JD → 调 Claude → 输出决策卡 → 人工决定投不投。

Returned to the simplest design: paste JD into terminal → call Claude → render decision card → human decides whether to apply.

结果：30 秒 / 岗位，零风控风险，可叠加台账留存与跨日复盘。这是目前在用的版本，承担了 108 个岗位的真实判断。

Outcome: 30 seconds per posting, zero risk-control exposure, with persistent logs for cross-day review. This is the version in production, having processed 108 actual postings so far.

心得：方案选型不是"哪个最先进"，是"哪个跟现实约束最匹配"。BOSS 平台高强度反爬就是约束，与其硬刚不如用最轻量的人机分工绕开它。

Takeaway: path selection is not about "which is most advanced" — it is about "which best matches the real constraints." Heavy anti-scraping is the constraint; rather than fighting it, use the lightest human-machine division to bypass it.

05 | 工作台 MVP 展示 MVP Demo

jd 命令行工具完整链路展示——从 JD 输入、AI 解析、决策卡输出到本地双轨留存。 / jd CLI tool full pipeline: from JD input through AI parsing to decision card output and local dual-track logging.

决策卡 CSV 台账（节选 20 行）——每条记录包含日期、公司、岗位、方向判断、核心要求、风险点、推荐简历版本、推荐招呼语版本、结论与投递理由。 / CSV decision log (excerpt of 20 records) — each entry captures date, company, role, direction analysis, key requirements, risk flags, recommended resume version, recommended outreach version, conclusion, and application reasoning.

06 | 方法论验证 Validation

当前数据 / Current Data（验证中 Validating）

累计判断岗位数：108
实际投递数：约 30–50 个（按"主投"和"可冲"档投递）
平均决策耗时：30 秒 / 岗位（手动判断对照组：约 5 分钟）
回复率：统计中（投递时间集中在五一假期，等待工作日反馈）

Effectiveness validation is incomplete by design — postings were sent during the May holiday and recruiter responses are expected after the holiday. The data section will be updated with reply rates and interview conversion in subsequent revisions.

初步观察 / Initial Observations

决策卡的"方向判断"在多次实测中有效识别了 title 与实际方向的偏差（例：title 标注"AI 产品经理（建筑专业）"，决策卡识别为"偏技术实验向，含 ComfyUI / 模型训练要求"）。

The "direction" field reliably caught title-vs-reality gaps in repeated tests.

简历版本推荐与个人事后判断的一致性较高，目前未发现明显错配。

Resume version recommendations have shown high consistency with retrospective judgment; no significant mismatch observed yet.

部分极端案例（要求 5 年经验 / 要求 985 名校 / 要求纯算法背景）会被决策卡明确标为"暂不主攻"，与人工判断一致——说明判断模板对"硬门槛"识别可靠。

Edge cases (5+ years required / Tier-1 universities only / pure algorithm background) are reliably flagged as "do not pursue" — confirming the template's reliability on hard filters.

07 | 反思与边界 Reflection & Boundary

这个工具能做什么、不能做什么，作为创作者需要写清楚。

What this tool can and cannot do — written plainly by its user.

它解决的事 / What It Does

把"识别同名异向、匹配简历版本、归类风险信号"这类结构化但重复的决策成本降到趋近于零；把使用者的精力从"读 100 份 JD"释放出来，转向"沉淀作品集 / 准备面试 / 学习行业知识"。

Drives the cost of structured-yet-repetitive decisions (catching title mismatches, matching resume versions, classifying risk signals) toward zero — and frees the user's energy from "read 100 JDs" to "build portfolio / prepare interviews / study the industry."

它没有解决的事 / What It Does Not Solve

转行简历被 HR 第一眼劝退的根本困境（这是平台展示机制 + 招聘市场结构问题，不是工具能解决的）
回复率本身（截至作品集发布日，108 个岗位的反馈仍在累积中）
面试转化率与 offer 质量（更长的反馈周期）

The structural reality that transition resumes get filtered fast (this is a platform display + hiring market issue, not a tooling issue), reply rate itself (data still accumulating), interview conversion and offer quality (longer feedback loop).

关于"半自动"的取舍 / On the "Semi-automated" Trade-off

工具刻意保留了"由使用者手动点击立即沟通 / 投递"这一步。

The tool deliberately keeps "click to chat / submit application" as a manual step performed by the user.

不是技术做不到自动投递，是：

自动投递触发平台风控（路径 1 已验证）
自动投递让使用者失去对每个岗位的最后一次审视机会
求职是高风险决策，保留人类最终决策权比追求"全自动"更重要

Not because automation is technically impossible, but because: auto-apply triggers platform risk-control (verified in Path 1); auto-apply removes the user's last review checkpoint per posting; job hunting is a high-stakes decision — preserving the human's final say matters more than chasing "fully automated."

这是工具的边界，也是它在 ops / 平台支持 / 客户成功这类岗位场景下应有的姿态——半自动不是"功能不完整"，是对人类判断权的有意保留。

This is the tool's boundary — and the posture it should maintain in ops / platform support / customer success contexts: "Semi-automated" is not "incomplete automation" — it is a deliberate preservation of human judgment.