hoelee-blog/docs/linkedin/05-stop-writing-agent-prompts-for-deterministic-work.md

LinkedIn Draft & Consultant Review: Stop Writing Agent Prompts for Deterministic Work

文档性质：从 blog.hoelee.com 改写的 LinkedIn 待发布技术动态。 发布状态：待顾问审阅 (未在 LinkedIn 发布) 源文章：https://blog.hoelee.com/posts/stop-writing-agent-prompts-for-deterministic-work/ 系统标识：stop-writing-agent-prompts-for-deterministic-work | 分类：notes 字数：1858 chars (LinkedIn 上限 3,000 chars)

---

一、 核心背景与铁律约束提示（供顾问审阅对照）

1. 求职画像契合（Senior 全栈/DevOps/AI 架构师）：严禁流露“做生意的老板/小公司创始人”倾向；一切商业实战均转换为“生产系统架构、高可用交付与自动化运维”视角。
2. 纯洁人设与零禁用词：绝对禁止 homelab, home lab, 家用, 小主机, 玩具, 折腾, 玩票 等词汇。必须使用 the infrastructure I manage, my self-hosted production nodes, production host。
3. 第一人称单数：统一使用单数 I / 我，杜绝虚张声势的复数 we。
4. 数据绝对保真：所有数字、性能比率、吞吐提升、故障现象严禁凭空捏造。
5. OPSEC 安全脱敏：无内部绝对路径（如 D:/dev），无真实私有内网 IP，公网 IP 必须转为文档保留地址。

---

二、 LinkedIn 拟发布动态文案

One of my automated background jobs failed 29 times in a row without sending an alert.

The job was trivial: query npm for the latest published version of a CLI tool, compare it with the currently installed version, and send a notification if a newer version exists.

Two strings and a version inequality. That was the entire task.

Yet I had initially architected it as an AI agent job, because that was the hammer in my hand at the time. Every cycle, an LLM parsed the script output and decided whether to alert. It ran fine until an upstream model routing change occurred. The agent framework recorded a model mismatch, triggered a silent drift skip, and exited cleanly without error.

The fix was not a more defensive prompt. It was deleting the LLM from the loop entirely.

The rule I enforce across all automated pipelines now:
**If the exit condition is a deterministic comparison — a string, a number, a status code — putting a model in that loop only adds failure modes.**

There is zero judgment required in `currentVersion < latestVersion`. Wrapping it in an LLM adds:
• Network latency and provider downtime
• Model drift and prompt deprecation risks
• Non-deterministic output formatting
• Continuous token billing

I replaced the agent loop with 125 lines of Python featuring explicit branch exit codes. It has run cleanly ever since.

Of the eleven automated operations loops I manage today, eight contain zero AI models in their execution path. The remaining three use LLMs exclusively where human ambiguity cannot be expressed as code: ranking unstructured emails, evaluating qualitative risk, and flagging stalled project contexts.

Where in your workflows are you using prompt engineering where a 10-line script would be infinitely more resilient?

Write-up: https://blog.hoelee.com/posts/stop-writing-agent-prompts-for-deterministic-work/

---

三、 外部顾问审阅评估反馈单

• 审阅结论：[ ] 通过 / [ ] 需微调 / [ ] 需重构 / [ ] 建议暂缓
• 1. 职场形象评估（是否符合高级技术雇佣人设，无老板/卖货感）：
• 2. 技术真实性与逻辑推导（因果链条与数据是否严谨）：
• 3. 语言地道度与人设去油（是否纯正地道、零业余玩具感）：
• 4. 优化修改建议（如有优化建议或修改文案，请在此批注）：