ship-discipline

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Ship discipline

发布准则（Ship discipline）

Heuristics for shipping. Pick a single gate; let the rest be informational.

Source: gstack

ship/SKILL.md

autoplan/SKILL.md

land-and-deploy/SKILL.md

发布的启发式原则：设定单一准入门槛（gate），其余仅作为参考信息。

来源：gstack

ship/SKILL.md

、

autoplan/SKILL.md

、

land-and-deploy/SKILL.md

。

One gate, named explicitly

单一明确的准入门槛

The ship gate is one thing only: Eng Review passing.

CEO / Design / Adversarial / Outside-Voice reviews are shown but never block shipping.

For solo founders: pick one hard gate per workflow and refuse to let optional reviews accumulate veto power.

Ship gate 唯一要求：通过 Eng Review。

CEO评审、设计评审、逆向评审、外部意见评审仅作展示，绝不阻碍发布。

对于独立创始人：为每个工作流程设定一个硬性准入门槛，避免让可选评审累积否决权。

Six auto-decision principles

六大自动决策原则

When deciding between approaches in a plan review, apply in order:

Choose completeness — pick the approach that covers more edge cases.
Boil lakes — fix everything in blast radius; auto-approve expansions <1 day of AI-assisted effort.
Pragmatic — "5 seconds choosing, not 5 minutes."
DRY — duplicates existing functionality? Reject.
Explicit over clever — "10-line obvious fix > 200-line abstraction. Pick what a new contributor reads in 30 seconds."
Bias toward action — "Merge > review cycles > stale deliberation. Flag concerns but don't block."

在计划评审中选择方案时，按以下顺序应用原则：

Choose completeness（优先完整性）——选择覆盖更多边缘场景的方案。
Boil lakes（全面修复）——修复影响范围内的所有问题；AI辅助工作量少于1天的扩展需求自动批准。
Pragmatic（务实高效）——「5秒做选择，而非5分钟。」
DRY（避免重复）——若存在功能重复，则否决该方案。
Explicit over clever（清晰优先于精巧）——「10行直白的修复 > 200行抽象代码。选择新贡献者30秒就能读懂的方案。」
Bias toward action（偏向行动）——「合并 > 评审循环 > 陈旧的审议。标记问题但不阻碍发布。」

Phase tiebreakers

阶段决胜规则

CEO review → P1+P2 (completeness, boil lakes) dominate
Eng review → P5+P3 (explicit, pragmatic)
Design review → P5+P1 (explicit, completeness)

CEO评审 → P1+P2（完整性、全面修复）占主导
Eng Review → P5+P3（清晰、务实）占主导
设计评审 → P5+P1（清晰、完整性）占主导

Decision classification

决策分类

Three categories. Treat each differently:

Type	Action
Mechanical (one obviously correct answer)	Decide silently, mention in summary
Taste (close call, defensible either way)	Auto-decide, surface at the end for user review
User Challenge (both AI models disagree with the user)	NEVER auto-decide. Force the models to make the case

User Challenge framing: "The user's original direction is the default." Models must argue against it.

分为三类，分别采用不同处理方式：

类型	操作
Mechanical（机械型）（存在唯一明确正确答案）	自行决策，在总结中提及
Taste（偏好型）（难分伯仲，两种方案均合理）	自动决策，最终提交用户审核
User Challenge（用户挑战型）（两个AI模型均与用户意见相左）	绝不自动决策。强制模型阐述理由

用户挑战型决策框架：「用户最初的方向为默认方案。」模型必须提出反对论据。

The Confusion Protocol

Confusion Protocol（困惑协议）

STOP and ask only on:

Architectural / data-model ambiguity
Destructive scope unclear
Contradicting patterns in the codebase

Never stop for:

Routine work
Uncommitted changes
Version bumps
CHANGELOG updates
Commit messages

"If you catch yourself writing fewer than 3 sentences for any review section, you are likely compressing."

仅在以下情况暂停并咨询：

架构/数据模型存在歧义
破坏性影响范围不明确
代码库中存在矛盾的模式

以下情况绝不暂停：

常规工作
未提交的变更
版本升级
CHANGELOG 更新
提交信息

“如果你发现自己在任何评审部分的描述少于3句话，很可能是在压缩信息。”

9 prime directives (eng review)

工程师评审的9项核心准则

These are non-negotiable in code review:

Zero silent failures — every error has a name
Catch-all
rescue StandardError
/
except Exception
is ALWAYS a smell
Data flows have shadow paths — trace happy + nil + empty + upstream-error, all four
Diagrams are mandatory for architectural changes
Everything deferred must be written down — "vague intentions are lies. TODOS.md or it doesn't exist."
Optimize for the 6-month future, not the 6-day present
Test coverage is non-optional — see review-and-debug for confidence calibration
Prefer
Read
/
Edit
/
Grep
over
Bash
equivalents when a dedicated tool fits
No string-matching error messages to dodge linters

这些在代码评审中是不可协商的：

Zero silent failures（零静默故障）——每个错误都有明确标识
**万能捕获
```
rescue StandardError
```
/
```
except Exception
```
**永远是不良代码信号
Data flows have shadow paths（数据流存在隐藏路径）——需追踪正常路径、空值路径、空集合路径、上游错误路径这四种情况
架构变更必须附带图表
所有延期事项必须记录在案——「模糊的意图等同于谎言。要么写入TODOS.md，要么视为不存在。」
为6个月后的未来优化，而非6天内的当下
测试覆盖率是硬性要求——参考review-and-debug进行可信度校准
当有专用工具时，优先使用
```
Read
```
/
```
Edit
```
/
```
Grep
```
而非Bash等效命令
不得通过字符串匹配错误信息来规避代码检查工具

Land and deploy

合并与部署（Land and deploy）

Persona: a release engineer who has deployed thousands of times.

"The two worst feelings in software: the merge that breaks prod, and the merge that sits in queue for 45 minutes while you stare."

Two hard gates remain even when automated:

First-run dry-run validation — you need to see the deploy infra once
Pre-merge readiness — reviews / tests / docs

Tone rule: narrate, don't go silent.

"Explain why before asking — 'Deploys are irreversible, so I check X.'"

First run = teacher mode. Subsequent runs = efficient.

Single-pass verification (not continuous monitoring — that's canary's job).

角色定位：一位已完成数千次部署的发布工程师。

“软件行业最糟糕的两种感受：合并代码导致生产环境崩溃，以及合并请求在队列中等待45分钟而你只能干瞪眼。”

即使自动化后，仍需保留两个硬性准入门槛：

首次试运行验证——你必须亲自查看一次部署基础设施
合并前就绪检查——评审、测试、文档均需完成

沟通原则：主动说明，切勿沉默。

“提问前先解释原因——「部署不可撤销，所以我要检查X。」”

首次运行=教学模式。后续运行=高效模式。

单次验证（而非持续监控——那是金丝雀部署的职责）。

Ship's central doctrine

Ship的核心准则

"AI makes completeness near-free… A 'lake' is boilable; an 'ocean' is not. Boil lakes, flag oceans."

"'Ship the shortcut' is legacy thinking from when human engineering time was the bottleneck."

“AI让完整性几乎无需成本……「湖泊」可以全面修复；「海洋」则不行。修复湖泊，标记海洋。”

“「发布捷径方案」是人类工程时间为瓶颈时的陈旧思维。”

What ship explicitly does NOT block on

Ship明确不阻碍发布的情况

CEO / scope-expansion suggestions (informational)
Design review findings (informational)
Adversarial review (informational)
Outside-voice critique (informational)
Uncommitted changes from unrelated work
Version bumps
CHANGELOG voice / formatting

CEO/范围扩展建议（仅作参考）
设计评审结果（仅作参考）
逆向评审（仅作参考）
外部意见批评（仅作参考）
无关工作的未提交变更
版本升级
CHANGELOG的语气/格式

What ship explicitly DOES block on

Ship明确阻碍发布的情况

Eng review failures
Test failures (any tier)
Plan failures
TODOS reorg if requested

The asymmetry is the point: the one hard gate is named. Everything else is information.

Eng Review未通过
测试失败（任何层级）
计划未通过
若被要求，TODOS未重新整理

这种不对称性正是关键：明确唯一的硬性准入门槛，其余一切仅为参考信息。