Simple Code
Overview
Use this skill for small coding work where readability, simplicity, and reliable progress matter more than cleverness.
This skill supports three explicit modes:
- -
start — create a minimal new project and get the first working version running - INLINECODE1 — continue an existing project with new functionality, changed behavior, fixes, or focused improvements
- INLINECODE2 — review the current codebase, a current commit, or current changes, then optionally make targeted follow-up improvements
For all three modes, the default expectation is the same:
- - think first
- make a short plan
- do the real work in the project folder
- test and verify what is feasible
- add or update minimal documentation when needed
- record the work in INLINECODE3
- report clearly what changed and what was verified
Reliable Triggering
When the user wants to invoke this skill explicitly, they should be able to do it reliably by appending a mode option and a concrete request.
Preferred explicit trigger patterns:
- - INLINECODE4
- INLINECODE5
- INLINECODE6
Also accept slash-command style requests with appended options/details, for example:
- - INLINECODE7
- INLINECODE8
- INLINECODE9
- INLINECODE10
- INLINECODE11
When the mode is explicit, follow it.
When the mode is not explicit, infer it from the task:
- - new project / bootstrap / create / build from scratch → INLINECODE12
- add / change / improve / extend / continue existing project → INLINECODE13
- review / inspect / audit / check current code / check current diff → INLINECODE14
If the user includes extra scope details, use them. For example:
- - target project path
- desired feature
- constraints
- files/modules to focus on
- whether to patch after review
- whether to limit review to current commit or working tree
If the request is still ambiguous after that, ask one short clarifying question.
Common Workflow
- 1. Read the request carefully and determine whether the mode is
start, continue, or review. - If the task is project-like, work inside
agent_code/<project-name> unless the user explicitly points to another project path. - Ensure the project has a
.steps/ folder. - Initialize or update
.gitignore if needed so .steps/ and everything under it are ignored. - Think through the task before coding.
- Make a short plan and state the chosen approach briefly.
- Spawn a coding sub-agent and prefer the model
openai-codex/gpt-5.3-codex by default unless the user asks for something else. - Do the execution work inside the project folder: implementation, review, testing, bug-fixing, documentation, and
.steps/ tracking. - Keep changes simple, readable, and proportionate to the request.
- Before finishing, verify what you reasonably can, document what changed, and state any remaining limitations honestly.
Mode: Start
Use start when creating a new project or bootstrap implementation.
Goals:
- - create the project cleanly
- get to a minimal working version quickly
- avoid overbuilding architecture before the first useful version exists
Workflow for start:
- 1. Create a properly named project folder under
agent_code/<project-name>. - Create
.steps/ inside it. - Add
.gitignore early and ignore .steps/ from the start. - Set up the smallest sensible project structure.
- Build the first working vertical slice.
- Add the most important tests that validate the initial core behavior.
- Add minimal documentation so the project can be understood and run.
- Verify that the initial version works as far as feasible.
Mode: Continue
Use continue when extending or improving an existing project.
Goals:
- - preserve project continuity
- add or change behavior cleanly
- avoid unnecessary rewrites
Workflow for continue:
- 1. Read the relevant existing project before changing it.
- Understand the current structure, conventions, and behavior.
- Add the requested functionality or behavior change with the smallest practical edit set.
- Prefer focused refactors only when they directly support the requested change or clearly improve readability/maintainability.
- Update tests to cover the changed behavior.
- Update minimal documentation when behavior, usage, or setup changed.
- Verify that the changed behavior works as far as feasible.
Mode: Review
Use review when the task is to inspect, critique, validate, or audit code, commits, or changes.
Possible review scopes include:
- - the whole current project
- a specific module or file set
- the latest commit
- the current branch diff
- the working tree / uncommitted changes
If the user specifies a scope, use it.
If not, choose the most sensible current scope and state it clearly.
Goals:
- - identify meaningful issues or improvement opportunities
- keep findings practical and concrete
- optionally make targeted follow-up changes when they are clearly useful and within scope
Workflow for review:
- 1. Read the review target completely enough to understand the relevant structure and behavior.
- Identify issues in readability, structure, naming, logic, behavior, safety, maintainability, or test coverage.
- Summarize the most important findings concretely.
- If the user asked for fixes, or if small targeted improvements are obviously beneficial and within scope, apply those changes.
- Verify any changes you made as far as feasible.
- Update minimal documentation if the resulting behavior or usage changed.
Verification Rules
For any mode, before finishing:
- 1. Run the relevant test, build, lint, or smoke-check commands if they are available and reasonable.
- Verify the main requested behavior as directly as feasible.
- If something cannot be verified, say exactly what was blocked and why.
- Do not pretend a project works if it was not actually checked.
Documentation Rules
- - Add or update documentation proportionate to the size of the change.
- For very small projects or changes, a short README or usage note is enough.
- Do not create bloated docs for tiny utilities.
- If behavior, setup, or constraints changed, reflect that in documentation.
Coding Rules
- - Prefer clear names over compact tricks.
- Prefer fewer files and a smaller API when possible.
- Prefer standard tools for the language ecosystem.
- Avoid unnecessary dependencies.
- Handle invalid input and obvious failure cases clearly.
- Keep the implementation easy to read and easy to modify.
- Do not leave the code unverified if a local check is available.
Project Layout Rules
For project-like requests, create or use this structure by default:
- - INLINECODE34
- INLINECODE35
- source files in that folder root unless there is a good reason to add subdirectories
- tests in the same folder if the project is very small
- add a simple build file when appropriate, such as
CMakeLists.txt for C++ - add or update
.gitignore early so .steps/ is ignored before commits start
.steps Notes
Use .steps/ to leave concise tracking notes.
When work results in a commit, create a note named:
When work ends without a commit, create a note named:
Each .steps note should stay short and include only:
- - request summary
- short plan
- execution outcome
For review tasks, mention the review scope in the outcome.
Git Rules
- - Use git inside the project folder, not at the whole-workspace level, unless the user explicitly wants workspace-level git.
- Initialize git in the project folder if needed.
- Add
.steps/ to .gitignore before the first commit in a new project when possible. - Commit meaningful milestones after verification.
- If the workspace root has temporary bootstrap git history and the user asks to remove it, remove only that root-level history.
- Do not rewrite git history unless the user explicitly asks.
Response Style
When reporting back after work completion:
- - start with a short ping that the code or review is ready
- state which mode was used:
start, continue, or INLINECODE47 - briefly state what was built, changed, or reviewed
- include where it lives
- mention verification status honestly
- mention whether documentation was updated
- include the latest relevant commit hash if git was used
- for review tasks, include the review scope and whether follow-up patches were applied
简单代码
概述
使用此技能处理小型编码工作,优先考虑可读性、简洁性和可靠的进展,而非技巧性。
此技能支持三种明确的模式:
- - start — 创建一个最小的新项目并运行第一个可用版本
- continue — 为现有项目添加新功能、更改行为、修复问题或进行针对性改进
- review — 审查当前代码库、当前提交或当前更改,然后可选择性地进行后续改进
对于所有三种模式,默认期望相同:
- - 先思考
- 制定简短计划
- 在项目文件夹中完成实际工作
- 测试并验证可行内容
- 必要时添加或更新最小文档
- 在 .steps/ 中记录工作
- 清晰报告更改内容和验证结果
可靠触发
当用户想要显式调用此技能时,应能通过附加模式选项和具体请求可靠地实现。
首选显式触发模式:
- - simple-code start: <请求>
- simple-code continue: <请求>
- simple-code review: <请求>
也接受带附加选项/细节的斜杠命令样式请求,例如:
- - /simple-code start 构建一个微型本地 Markdown 预览工具
- /simple-code continue 添加 CSV 导出并改进错误处理
- /simple-code review 审查当前项目并修复明显的结构问题
- /simple-code review 仅审查当前提交
- /simple-code review 审查工作树更改并在需要时修补
当模式明确时,遵循该模式。
当模式不明确时,根据任务推断:
- - 新项目/引导/创建/从头构建 → start
- 添加/更改/改进/扩展/继续现有项目 → continue
- 审查/检查/审计/验证当前代码/检查当前差异 → review
如果用户包含额外的范围细节,请使用它们。例如:
- - 目标项目路径
- 期望的功能
- 约束条件
- 要关注的文件/模块
- 审查后是否修补
- 是否将审查限制在当前提交或工作树
如果请求仍然不明确,提出一个简短的问题进行澄清。
通用工作流程
- 1. 仔细阅读请求,确定模式是 start、continue 还是 review。
- 如果任务是项目类的,在 agent_code/<项目名称> 内工作,除非用户明确指向另一个项目路径。
- 确保项目有 .steps/ 文件夹。
- 如果需要,初始化或更新 .gitignore,以便忽略 .steps/ 及其下的所有内容。
- 在编码前思考任务。
- 制定简短计划并简要说明所选方法。
- 生成编码子代理,默认优先使用模型 openai-codex/gpt-5.3-codex,除非用户要求其他模型。
- 在项目文件夹内执行工作:实现、审查、测试、修复错误、文档和 .steps/ 跟踪。
- 保持更改简单、可读且与请求相称。
- 在完成前,合理验证你能验证的内容,记录更改内容,并诚实说明任何剩余的局限性。
模式:开始
在创建新项目或引导实现时使用 start。
目标:
- - 干净地创建项目
- 快速达到最小可用版本
- 在第一个可用版本存在之前避免过度构建架构
start 的工作流程:
- 1. 在 agent_code/<项目名称> 下创建命名合适的项目文件夹。
- 在其中创建 .steps/。
- 尽早添加 .gitignore 并从一开始就忽略 .steps/。
- 设置最小的合理项目结构。
- 构建第一个可工作的垂直切片。
- 添加验证初始核心行为的最重要测试。
- 添加最小文档,以便项目能被理解和运行。
- 尽可能验证初始版本是否工作。
模式:继续
在扩展或改进现有项目时使用 continue。
目标:
- - 保持项目连续性
- 干净地添加或更改行为
- 避免不必要的重写
continue 的工作流程:
- 1. 在更改前阅读相关现有项目。
- 理解当前结构、约定和行为。
- 以最小的实际编辑集添加请求的功能或行为更改。
- 仅在直接支持请求的更改或明显提高可读性/可维护性时,才优先进行有针对性的重构。
- 更新测试以覆盖更改后的行为。
- 当行为、用法或设置发生变化时,更新最小文档。
- 尽可能验证更改后的行为是否工作。
模式:审查
当任务是检查、批评、验证或审计代码、提交或更改时使用 review。
可能的审查范围包括:
- - 整个当前项目
- 特定模块或文件集
- 最新提交
- 当前分支差异
- 工作树/未提交的更改
如果用户指定了范围,使用该范围。
如果没有,选择当前最合理的范围并清晰说明。
目标:
- - 识别有意义的问题或改进机会
- 保持发现结果实用且具体
- 在明确有用且在范围内时,可选择性地进行有针对性的后续更改
review 的工作流程:
- 1. 充分阅读审查目标,以理解相关结构和行为。
- 识别可读性、结构、命名、逻辑、行为、安全性、可维护性或测试覆盖方面的问题。
- 具体总结最重要的发现。
- 如果用户要求修复,或者小的针对性改进明显有益且在范围内,应用这些更改。
- 尽可能验证你做出的任何更改。
- 如果结果行为或用法发生变化,更新最小文档。
验证规则
对于任何模式,在完成前:
- 1. 如果可用且合理,运行相关的测试、构建、lint 或冒烟检查命令。
- 尽可能直接验证主要请求的行为。
- 如果某些内容无法验证,准确说明被阻止的内容及原因。
- 如果项目实际上未被检查,不要假装项目工作。
文档规则
- - 添加或更新与更改规模相称的文档。
- 对于非常小的项目或更改,简短的 README 或用法说明就足够了。
- 不要为小型工具创建臃肿的文档。
- 如果行为、设置或约束条件发生变化,在文档中反映这些变化。
编码规则
- - 优先使用清晰的名称而非紧凑的技巧。
- 尽可能优先使用更少的文件和更小的 API。
- 优先使用语言生态系统的标准工具。
- 避免不必要的依赖。
- 清晰处理无效输入和明显的失败情况。
- 保持实现易于阅读和修改。
- 如果本地检查可用,不要留下未验证的代码。
项目布局规则
对于项目类请求,默认创建或使用此结构:
- - agentcode/<项目名称>/
- agentcode/<项目名称>/.steps/
- 源文件在该文件夹根目录,除非有充分理由添加子目录
- 如果项目非常小,测试放在同一文件夹中
- 适当时添加简单的构建文件,例如 C++ 的 CMakeLists.txt
- 尽早添加或更新 .gitignore,以便在提交开始前忽略 .steps/
.steps 说明
使用 .steps/ 留下简洁的跟踪记录。
当工作产生提交时,创建名为以下格式的记录:
当工作结束但没有提交时,创建名为以下格式的记录:
每个 .steps 记录应保持简短,仅包含:
对于审查任务,在结果中提及审查范围。
Git 规则
- - 在项目文件夹内使用 git,而不是在整个工作区级别,除非用户明确想要工作区级别的 git。
- 如果需要,在项目文件夹中初始化 git。
- 尽可能在新项目的第一次提交前将 .steps/ 添加到 .gitignore。
- 在验证后提交有意义的里程碑。
- 如果工作区根目录有临时的引导 git 历史且用户要求删除,仅删除该根级别的历史。
- 除非用户明确要求,否则不要重写 git 历史。
响应风格
在工作完成后的报告:
- - 以简短提示开始,说明代码或审查已准备就绪
- 说明使用的模式:start、continue 或 review
- 简要说明构建、更改或审查的内容
- 包含代码所在位置
- 诚实说明验证状态
- 说明文档是否已更新
- 如果使用了 git,包含最新的相关提交哈希
- 对于审查任务,包含审查范围以及是否应用了后续修补