马具创造者
- 作者仓库星标 7,591
- 作者仓库 learn-harness-engineering
- 领域
- 通用
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @walkinglabs · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 即装即用
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- macOS · Linux · Windows
- 运行环境要求
- Node.js
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
# Harness Creator
Use this skill to make a repository easier for coding agents to start, stay in scope, verify work, and resume across sessions. Keep the harness small enough that agents actually follow it.
Not for model selection, prompt tuning in isolation, chat UI design, or general app architecture.
## Core Model
Every useful coding-agent harness has five subsystems:
| Subsystem | Minimal artifact | Purpose |
|---|---|---|
| Instructions | `AGENTS.md` or `CLAUDE.md` | Startup path, working rules, definition of done |
| State | `feature_list.json`, `progress.md` | Current feature, status, evidence, next step |
| Verification | `init.sh` or documented commands | Tests/checks the agent must run before claiming done |
| Scope | Feature dependencies and done criteria | Prevents overreach and half-finished work |
| Lifecycle | `session-handoff.md`, end-of-session routine | Makes the next session restartable |
## First Move
1. Inspect what already exists: instruction files, feature/state files, verification commands, docs, package manifests.
2. Ask only for missing context that cannot be inferred safely: target agent, desired file name, tolerance for structure, and whether overwriting is allowed.
3. Prefer a minimal harness first. Add memory, tool safety, multi-agent, or benchmark details only when the user's problem calls for them.
## Common Tasks
### Create a harness
Use the bundled script when working on a local repository:
```bash
node skills/harness-creator/scripts/create-harness.mjs --target /path/to/project
```
Options:
- `--agent-file CLAUDE.md` for Claude-oriented projects.
- `--package-manager npm|pnpm|yarn|bun` when detection is wrong.
- `--commands "cmd one,cmd two"` for custom verification.
- `--force` only after confirming overwrites are acceptable.
Then explain what was created and how the user should replace placeholder feature entries.
### Audit an existing harness
Run:
```bash
node skills/harness-creator/scripts/validate-harness.mjs --target /path/to/project
```
Report the five subsystem scores, the lowest-scoring area, and the first 2-3 changes that would improve reliability. Treat the lowest score as a candidate bottleneck; confirm with failures, logs, or task outcomes before claiming causality.
### Produce a report
Use when the user wants a shareable assessment:
```bash
node skills/harness-creator/scripts/render-assessment-html.mjs --target /path/to/project
node skills/harness-creator/scripts/run-benchmark.mjs --target /path/to/project --html /path/to/report.html
```
Be clear that this is a structural benchmark. Real effectiveness still needs before/after agent sessions on representative tasks.
## When to Read References
Load only the reference needed for the user's problem:
- Memory across sessions: [Memory Persistence](references/memory-persistence-pattern.md)
- Reusable workflows as skills: [Skill Runtime](references/skill-runtime-pattern.md)
- Permissions, tools, concurrency: [Tool Registry & Safety](references/tool-registry-pattern.md)
- Context budget and progressive disclosure: [Context Engineering](references/context-engineering-pattern.md)
- Delegation and parallel agents: [Multi-Agent Coordination](references/multi-agent-pattern.md)
- Hooks, startup, long-running work: [Lifecycle & Bootstrap](references/lifecycle-bootstrap-pattern.md)
- Non-obvious failure modes: [Gotchas](references/gotchas.md)
## Design Rules
- Keep the root instruction file short: routing and invariants, not a full manual.
- Put project facts in project docs, not in the skill.
- Make verification commands explicit and runnable.
- Require evidence before marking a feature done.
- Use one active feature unless the harness has explicit multi-agent ownership boundaries.
- Prefer append/update state files over relying on chat history.
- Never hide destructive behavior in scripts; overwrites require explicit user approval.
## Deliverable Checklist
For a usable minimal harness, leave the target project with:
- [ ] `AGENTS.md` or `CLAUDE.md`
- [ ] `feature_list.json`
- [ ] `progress.md`
- [ ] `init.sh`
- [ ] Optional `session-handoff.md` for multi-session work
- [ ] Documented verification evidence or next action
If you cannot create files, provide exact file contents and commands instead.
章节与文件索引
按说明文件中的顺序排列。
核心模型
第一步
创建线束
审核现有线束
制作报告
文件/命令 -> AGENTS.md · CLAUDE.md · feature_list.json · progress.md · init.sh · session-handoff.md · --agent-file CLAUDE.md · --package-manager npm|pnpm|yarn|bun
内容校验值 -> 2bb35f0c2f43
当前原文的自动摘译,尚未人工审校。
核心模型
每个有用的编码代理工具都有五个子系统: 子系统·最小工件·目的说明·
AGENTS.md或CLAUDE.md·启动路径、工作规则、完成状态的定义·feature_list.json、progress.md·当前功能、状态、证据、下一步验证·init.sh或记录的命令·在声明完成之前测试/检查代理必须运行第一步
检查已存在的内容:指令文件、功能/状态文件、验证命令、文档、包清单。仅询问无法安全推断的缺失上下文:目标代理、所需的文件名、结构容差以及是否允许覆盖。首先选择最小的安全带。仅当用户的问题需要时才添加内存、工具安全、多代理或基准测试详细信息。
创建线束
在本地存储库上工作时使用捆绑的脚本:选项:
--agent-file CLAUDE.md适用于面向 Claude 的项目。--package-manager npm|pnpm|yarn|bun检测错误时。--commands "cmd one,cmd two"用于自定义验证。审核现有线束
运行:报告五个子系统得分、得分最低的区域以及可提高可靠性的前 2-3 个更改。将最低分数视为候选瓶颈;在声明因果关系之前先确认故障、日志或任务结果。
制作报告
当用户想要可共享的评估时使用:请明确这是一个结构基准。真正的有效性仍然需要座席在代表任务上进行会议之前/之后。