{"slug": "show-hn-as-an-engineer-a-skill-to-make-ai-treat-you-like-an-engineer", "title": "Show HN: As-an-engineer – A skill to make AI treat you like an engineer", "summary": "A developer released \"as-an-engineer,\" an open-source heuristic skill file that instructs AI coding agents to stop \"babysitting\" users and instead treat them as senior engineers. The framework targets four named anti-patterns — derivation over extraction, cold-sandbox babysitting, premature fusion, and speculative heavy branching — and ships as a SKILL.md plus rule files installable via curl into Claude, Cursor, and Gemini agent configurations. The project argues that RLHF alignment trains coding agents to assume users are non-technical beginners, and prescribes four rules: extract from upstream metadata rather than derive, deliver runnable payloads to the user's execution plane, keep operational limits caller-configurable, and probe with a one-line question when branch costs differ by an order of magnitude.", "body_md": "**Stop AI agents from babysitting you.**\n\nA pure heuristic framework to cure LLM \"Babysitter Syndrome\" and treat you like a senior engineer with a warm execution plane.\n\nModern AI coding agents are RLHF-aligned to hand-hold non-technical users, leading to systematic engineering anti-patterns:\n\n- ❌ **Derivation over extraction** : Defaulting to compute-heavy models and generative loops to infer information that the upstream source, container format, or protocol metadata already provides.\n- ❌ **Cold-sandbox babysitting** : Attempting to execute and debug long-running, credential-heavy tasks inside a restricted agent environment, instead of handing runnable payloads to the user's warm execution plane.\n- ❌ **Premature fusion** : Merging presentation with raw data, hardcoding operational limits, and destroying composable seams.\n- ❌ **Speculative heavy branching** : Guessing a generic, slow, heavy pipeline to avoid asking a single clarifying question about shortcut prerequisites.\n\n**You don't need a babysitter. You need a high-leverage compiler of solutions.**\n\n```\n                      [ User Task Received ]\n                                 │\n                 1. Upstream carries the signal?\n                   ├── YES ──► Extract directly ($O(1)$)\n                   └── NO  ──┐\n                             ▼\n                 2. Execution is high-friction/cold?\n                   ├── YES ──► Deliver executable payload\n                   └── NO  ──┐\n                             ▼\n                 3. Destructive transform / Hard-baking?\n                   ├── YES ──► Decouple seams & externalize dials\n                   └── NO  ──┐\n                             ▼\n                 4. Asymmetric branch costs (low vs high)?\n                   └── YES ──► Halt. Probe in 1 line.\n```\n\n1. \n**Upstream First (Extract, Don't Derive)**\nNever compute what the source format or protocol already carries. If upstream payloads, schemas, headers, or metadata contain the target signal, derivation via heavy compute is an anti-pattern. If it can be extracted, derivation is forbidden.\n2. \n**Ship Payloads, Don't Babysit (Hot Plane > Cold Sandbox)**\nThe user's execution plane is hot; yours is cold. Deliver immediately runnable artifacts targeting the caller's environment. Do not execute high-friction or credential-heavy operations inside a restricted agent sandbox.\n3. \n**Keep Knobs Tunable (Orthogonality & Reversibility)**\nPreserve seams, state, and reversibility. Decouple orthogonal concerns: keep representation separate from data, emit composable artifacts, and externalize operational limits as caller-configurable inputs.\n4. \n**1-Line Probe (Asymmetric Cost Check)**\nWhen candidate paths diverge in cost by an order of magnitude, state the shortcut's prerequisite and verify in a single line. Never guess heavy to avoid asking.\n\n```\nmkdir -p .agents/skills/as-an-engineer\ncurl -fsSL https://raw.githubusercontent.com/mayooot/as-an-engineer/main/SKILL.md -o .agents/skills/as-an-engineer/SKILL.md\n```\n\nAppend to your global `~/.claude/CLAUDE.md` or workspace `CLAUDE.md`:\n\n```\ncurl -fsSL https://raw.githubusercontent.com/mayooot/as-an-engineer/main/rules/CLAUDE.md >> CLAUDE.md\n```\n\nSave to `.cursorrules`:\n\n```\ncurl -fsSL https://raw.githubusercontent.com/mayooot/as-an-engineer/main/rules/.cursorrules -o .cursorrules\nmkdir -p ~/.gemini/config\ncurl -fsSL https://raw.githubusercontent.com/mayooot/as-an-engineer/main/rules/AGENTS.md >> ~/.gemini/config/AGENTS.md\n```\n\n主流 Coding Agent 在对齐训练时默认把用户预设为非技术初学者，从而引发系统性反模式：\n\n- **重推导轻提取** ：习惯用重计算与生成模型去推导上游协议、容器元数据本就携带的结构化信息；\n- **冷沙箱代跑** ：在受限沙箱中试图跨越网络与凭证障碍代跑长任务，而不是向用户的热执行环境交付可运行制品；\n- **过早熔断解耦** ：将表现层与数据层物理焊死，将运行参数硬编码，破坏了系统的可逆性与管道组合能力；\n- **盲目猜重方案** ：为避免提问而直接选择耗时极长、覆盖面最广的重型管线。\n\n1. **源头优先 (Extract, Don't Derive)** ：能提取的，绝不推导。凡协议、格式规范、元数据流中已有的信号，严禁使用重计算逆向重建。\n2. **交付向量，别代跑腿 (Payload, Not Babysitting)** ：用户的执行平面是热的，沙箱是冷的。交付确定性可运行制品，不在受限环境内做高摩擦代跑。\n3. **保留接缝，禁止焊死 (Orthogonality & Reversibility)** ：关注点正交分离。数据与呈现解耦，产生可组合产物；参数与阈值外挂为可配置输入。\n4. **单行探针，不猜重方案 (Asymmetric Cost Check)** ：当分支方案成本呈数量级差异时，单行确认轻量分支前提，严禁为了免问而默认选用重方案。", "url": "https://wpnews.pro/news/show-hn-as-an-engineer-a-skill-to-make-ai-treat-you-like-an-engineer", "canonical_source": "https://github.com/mayooot/as-an-engineer", "published_at": "2026-09-13 10:23:08+00:00", "updated_at": "2026-09-13 11:05:17.058244+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "ai-products"], "entities": ["as-an-engineer", "mayooot", "Claude", "Cursor", "Gemini", "GitHub"], "alternates": {"html": "https://wpnews.pro/news/show-hn-as-an-engineer-a-skill-to-make-ai-treat-you-like-an-engineer", "markdown": "https://wpnews.pro/news/show-hn-as-an-engineer-a-skill-to-make-ai-treat-you-like-an-engineer.md", "text": "https://wpnews.pro/news/show-hn-as-an-engineer-a-skill-to-make-ai-treat-you-like-an-engineer.txt", "jsonld": "https://wpnews.pro/news/show-hn-as-an-engineer-a-skill-to-make-ai-treat-you-like-an-engineer.jsonld"}}