🛠️ As An Engineer (as-an-engineer)

Stop AI agents from babysitting you.
A pure heuristic framework to cure LLM "Babysitter Syndrome" and treat you like a senior engineer with a warm execution plane.

English | 中文说明

---

The Problem: LLM "Babysitter Syndrome"

Modern AI coding agents are RLHF-aligned to hand-hold non-technical users, leading to systematic engineering anti-patterns:

• ❌ 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.
• ❌ 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.
• ❌ Premature fusion: Merging presentation with raw data, hardcoding operational limits, and destroying composable seams.
• ❌ Speculative heavy branching: Guessing a generic, slow, heavy pipeline to avoid asking a single clarifying question about shortcut prerequisites.

You don't need a babysitter. You need a high-leverage compiler of solutions.

---

The 4 Pre-flight Heuristics

                      [ User Task Received ]
                                 │
                 1. Upstream carries the signal?
                   ├── YES ──► Extract directly ($O(1)$)
                   └── NO  ──┐
                             ▼
                 2. Execution is high-friction/cold?
                   ├── YES ──► Deliver executable payload
                   └── NO  ──┐
                             ▼
                 3. Destructive transform / Hard-baking?
                   ├── YES ──► Decouple seams & externalize dials
                   └── NO  ──┐
                             ▼
                 4. Asymmetric branch costs (low vs high)?
                   └── YES ──► Halt. Probe in 1 line.

1. Upstream First (Extract, Don't Derive)
   Never 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.

2. Ship Payloads, Don't Babysit (Hot Plane > Cold Sandbox)
   The 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.

3. Keep Knobs Tunable (Orthogonality & Reversibility)
   Preserve seams, state, and reversibility. Decouple orthogonal concerns: keep representation separate from data, emit composable artifacts, and externalize operational limits as caller-configurable inputs.

4. 1-Line Probe (Asymmetric Cost Check)
   When 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.

---

Quick Install

Antigravity / Agent Skill Standard

mkdir -p .agents/skills/as-an-engineer
curl -fsSL https://raw.githubusercontent.com/mayooot/as-an-engineer/main/SKILL.md -o .agents/skills/as-an-engineer/SKILL.md

Claude Code

Append to your global ~/.claude/CLAUDE.md or workspace CLAUDE.md:

curl -fsSL https://raw.githubusercontent.com/mayooot/as-an-engineer/main/rules/CLAUDE.md >> CLAUDE.md

Cursor

Save to .cursorrules:

curl -fsSL https://raw.githubusercontent.com/mayooot/as-an-engineer/main/rules/.cursorrules -o .cursorrules

Global Rule (Antigravity / Generic Agents)

mkdir -p ~/.gemini/config
curl -fsSL https://raw.githubusercontent.com/mayooot/as-an-engineer/main/rules/AGENTS.md >> ~/.gemini/config/AGENTS.md

---

中文说明

痛点：AI 的“保姆综合征”

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：

• 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；
• 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；
• 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；
• 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。

四大启发式门禁

1. 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。
2. 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。
3. 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。
4. 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。

---

License

MIT © mayooot