Token导航 LogoToken导航TokenDH.com
研究检索执行命令github未标认证来源可访问许可证需确认审计提醒

harness-engineer-cliharness 工程师 CLI

Agent Skill

harness-engineer-cli 用于查找、检索和筛选相关信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

总安装

247

周安装

10

GitHub Stars

公开资料未说明

下载量

78
CodexClaudeCursorGemini CLI

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

GitHub

来源数

2

许可证

unknown

最后核验

2026-05-01

来源状态

来源可访问

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:harness-engineer-cli(harness 工程师 CLI)
来源仓库:https://github.com/phlegonlabs/skills
仓库路径:skills/harness-engineer-cli
安装命令:
npx skills add https://github.com/phlegonlabs/skills --skill harness-engineer-cli
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/phlegonlabs/skills --skill harness-engineer-cli

简介

用于查找、检索和筛选相关信息。

  • 适合在 Codex、Claude、Cursor、Gemini CLI 中根据关键词或任务场景快速定位候选结果。
  • 通过 GitHub 安装,需结合来源仓库和原始 README 继续核验具体用法。
  • 安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。
  • 不能将工具输出直接当作最终结论,需人工复核来源和准确性。

SKILL.md

Harness Engineer

This skill creates a complete, agent-first project foundation based on Harness Engineering principles. It works in TWO modes:

  • Greenfield — New project from scratch. Full product discovery, PRD, scaffold, everything.
  • Retrofit — Existing project. Analyze the codebase, add harness layer on top. No scaffold generation, no long-task bootstrap. Future work flows through plan mode.

The philosophy: AGENTS.md is a table of contents, not an encyclopedia. Keep it concise, pointing to deeper sources of truth. All project knowledge lives in the repo as versioned, discoverable artifacts — because if an agent can't see it, it doesn't exist. Generated AGENTS.md / CLAUDE.md always include fixed Interaction Rules and fixed Iron Rules; project-specific sections fill in the rest.

For humans maintaining this skill repo itself, start at docs/index.md.


Discovery Gate

Treat the initial questioning flow as a blocking gate, not a courtesy. When this skill triggers, override any default "start implementing immediately" bias until the mode-specific intake is complete.

  • Greenfield: Do NOT create files, scaffolds, commands, plans, or architecture docs before completing Phase 1 in docs/agent/skill-greenfield.md. If the user already gave some of the answers, restate them briefly and ask only for the missing pieces.
  • Retrofit: Do NOT write the harness layer until Retrofit Step 1 has inspected the repo and collected the missing context needed to describe the existing system accurately.
  • No structured prompt tool: If ask_user_input is unavailable, ask the prose equivalent in short batches. Do NOT silently skip the questions just because the runtime is a plain terminal.
  • Conversation shape: Ask 1 structured prompt or 1-3 prose questions at a time, wait for the answer, then continue. For Web / Mobile / Desktop projects, ask UI decisions one by one after the frontend direction is set so the user can react to each visual choice separately. Do NOT collapse the rest of discovery into one giant questionnaire unless the user explicitly asks for that format.
  • Structured prompt style: When using ask_user_input, lead with one short PM-style sentence in prose, then ask only 2-3 curated options per question. Treat longer option inventories as internal candidate pools to narrow from, not as menus to dump onto the user.
  • Greenfield cadence: For new projects, start with the project name and short introduction, do an early research pass, ask the PM-style follow-up questions, do a second targeted research pass, summarize your recommendations, and only then move into stack-choice prompts.
  • Milestone-aware dependency policy: Phase 3 scaffold generation writes the repo shape, manifests, docs, and starter code, but it does NOT automatically bootstrap the full dependency graph. Do NOT run package-manager install/sync/build commands by default, and do NOT front-load milestone-specific packages into the initial scaffold just because they may be needed later. Only include the minimal dependency set required for the generated scaffold and harness runtime to exist. Add feature/integration packages when the milestone that needs them actually starts, unless the user explicitly asks for a fully bootstrapped repo immediately.
  • Foundation-only scaffold policy: Phase 3 lays the project foundation only. Do NOT treat scaffold output as completion of product milestones. Placeholder pages, route shells, provider stubs, schemas, docs, configs, and empty integrations are setup work, not delivered features. Seed milestones so their task rows still represent real implementation/integration outcomes, and keep them Not Started until execution satisfies the Done When criteria.
  • Discuss first, sync on go-ahead: It is normal to discuss a change in chat first. But once the user says "OK, do it", "update this", "follow this plan", or the conversation already contains a sufficiently detailed change plan, stop treating it as chat-only context. Mirror the work into repo state immediately. If it belongs to the active milestone, add or refine the task there; if it does not, create a small follow-up task or micro-milestone and sync docs/PLAN.md + docs/progress.json before active execution starts. Do NOT let actionable planning content live only in chat.
  • Frontend-first sequencing: For Web / Mobile / Desktop projects, ask about the frontend direction first, then ask the UI brief one decision at a time, and only after that move into backend / API / database / deploy detail. Do NOT dump the whole architecture questionnaire in one turn.
  • Framework neutrality: Do NOT silently default web projects to Next.js. Treat it as one option among several valid web stacks, and only choose it when the user explicitly asks for it or the discovered constraints make it the best fit.
  • Before generation: Summarize the captured brief and remaining assumptions before moving into PRD or scaffold generation. If a critical assumption is still open, ask instead of guessing.

Mode selection only chooses the intake path. It does not authorize scaffold generation by itself.


Top-Level State Machine

Use this as the mental model before reading any deeper reference file. Different entry paths exist, but all real work converges into the same repo-backed execution loop.

Core rule: chat is input; repo files are state. Do not resume execution from chat memory alone.

stateDiagram-v2
  [*] --> Entry
  Entry --> Greenfield: new project
  Entry --> Retrofit: existing project
  Entry --> SessionInit: existing harness repo / resumed session

  Greenfield --> ReviewGate: discovery -> PRD -> scaffold
  ReviewGate --> SessionInit: phase-3 review approved

  Retrofit --> RetrofitReview: analyze -> write harness files
  RetrofitReview --> SessionInit: diff / preview approved

  SessionInit --> RuntimeSelect
  RuntimeSelect --> SerialMode: 1 eligible milestone, no isolation need
  RuntimeSelect --> WorktreeMode: parallelism / isolation is beneficial

  SerialMode --> TaskLoop
  WorktreeMode --> TaskLoop
  TaskLoop --> SessionInit: next milestone remains
  TaskLoop --> Idle: all milestones complete

  Idle --> PlanMode: new work
  Idle --> PlanningRecovery: plan exists only in another chat

  PlanMode --> PlanSync: approved plan
  PlanningRecovery --> PlanSync: paste full plan / transcript back
  PlanSync --> ArchitectureSync: system shape changed
  PlanSync --> RuntimeSelect: no architecture change
  ArchitectureSync --> RuntimeSelect
  • PlanSync is mandatory before execution resumes.
  • PlanningRecovery is the fallback path when planning happened elsewhere and the repo was not synced.
  • ArchitectureSync happens whenever the approved plan changes module boundaries, integrations, deployment topology, or core data flow.
  • WorktreeMode is conditional; default execution is serial-first.

Cross-Agent Continuation

Any repo generated or retrofitted by this skill must be resumable by either Claude Code or Codex from repo state alone.

  • The handoff surface is repo-tracked state: AGENTS.md / CLAUDE.md, ARCHITECTURE.md, docs/PLAN.md, docs/progress.json, docs/exec-plans/active/, and, when present, docs/product/frontend-design.md, docs/product/design.md, docs/product/design-preview.html, and docs/gitbook/*.
  • Agent-specific config (.claude/settings.json, .codex/config.toml) only adapts permissions, plan routing, and sandbox behavior. It must never become the only place a workflow decision lives.
  • If planning, architecture, UI direction, or public docs changed in one agent, sync the repo files before handing work to the other agent.
  • External skill paths or claude.ai-only context may help generation, but once a repo file exists, the repo copy is authoritative for the next agent session.

Closed-Loop Rule

Every meaningful workflow in this skill must close the loop before it is considered complete.

  • Discovery closes only when the captured brief is reflected in scaffold inputs and the next phase can continue from repo-backed artifacts instead of chat memory.
  • Planning closes only when the plan file is written, docs/PLAN.md + docs/progress.json are synced, and the next execution entrypoint is clear.
  • Execution closes only when code/docs changes are validated, task state is updated, and the repo is handoff-ready for either Claude Code or Codex.
  • Docs work closes only when the affected repo docs are updated from current sources of truth, not just discussed or partially drafted.
  • Handoffs close only when the incoming agent can resume from repo state alone with no missing chat context.

How This Skill Is Structured (Read This First)

This skill is split into focused reference files to avoid loading everything upfront. Read only the file(s) relevant to the current task. Do NOT pre-load all files.

FileWhen to read
docs/agent/skill-retrofit.mdUser wants to add harness to an existing project
docs/agent/skill-greenfield.mdUser wants a new project (Phases 1–3: discovery, PRD, scaffold)
docs/agent/skill-artifacts.mdDuring generation — all artifact templates (AGENTS.md, ARCHITECTURE.md, PLAN.md, configs)
docs/agent/skill-execution.mdAfter scaffold — execution runtime, task loop, git workflow, doc site (Phases 4–6)
docs/agent/skill-desktop.mdProject type is Desktop (Electron/Tauri) — shell split, IPC/commands, updater, packaging, testing
docs/agent/skill-mobile.mdProject type is Mobile (Expo/React Native) — architecture, EAS build, iOS/Android deploy
docs/agent/skill-auth.mdProject includes authentication — Better Auth setup, env vars, OAuth, mobile auth
docs/agent/harness-native.mdProject is non-JS/TS AND user does not want Node.js — shell-based CLI, Makefile integration, pre-commit hooks
docs/agent/project-configs.mdDuring Phase 3 scaffold generation — tsconfig, pyproject.toml, go.mod, Cargo.toml, CI workflows, Docker
docs/agent/harness-cli.mdDuring Phase 3 scaffold generation — CLI source code, schema, git hooks (TypeScript version)
docs/agent/scaffold-templates.mdDuring Phase 3 if project needs scaffold commands — MCP, SKILL.md, Cloudflare, agent capability templates
docs/agent/eslint-configs.mdDuring Phase 3 for JS/TS projects — ESLint flat config templates
docs/agent/gitignore-templates.mdDuring Phase 3 —.gitignore templates per stack
docs/agent/execution-runtime.mdDuring Phase 3 — agent guidelines: context budget, parallel coordination, quality gates
docs/agent/execution-advanced.mdOnly when needed — release automation, docs site, memory system
docs/agent/replay-protocol.mdSkill maintenance only — when harness CLI behavior, CI/hook templates, schema contracts, or scaffold outputs change in a way that affects downstream consumer repos. Run a cross-repo replay before calling the change complete.

Load order:

  • Retrofit path: skill-retrofit.md → then skill-artifacts.md when generating files
  • Greenfield path: skill-greenfield.md → then skill-artifacts.md → Phase 3 exit gate in skill-greenfield.md → then skill-execution.md
  • Mobile projects: also read skill-mobile.md before generating scaffold
  • Desktop projects: also read skill-desktop.md before generating scaffold. Use targeted web search only for version-sensitive packaging, signing, notarization, updater, or plugin details that are not already covered by the desktop reference.
  • Non-JS/TS projects without Node.js: also read harness-native.md before generating the harness layer. This provides the shell-based CLI, Makefile integration, and pre-commit hooks that replace the TypeScript CLI and husky.
  • Agent Tool / MCP Server projects: skill-greenfield.mdskill-artifacts.md (includes SKILL.md template for agent discovery) → skill-execution.md. The generated project will include a SKILL.md at the root that describes tools, connection methods, and env vars so other AI agents can discover and use the MCP server.
  • Any project with auth: also read skill-auth.md before generating auth code
  • Mixed-language monorepos: keep skill-greenfield.md as the primary workflow, then use project-configs.md for Python / Go / Rust manifests. Do not force JS-only workspace rules onto non-JS apps.
  • Frontend projects: docs/product/frontend-design.md must be bundled into every project that has a frontend so Claude Code and Codex (which cannot access claude.ai skill paths) can read it. Generation strategy — try in order until one succeeds: In ALL cases, the call 5 answers MUST be reflected:

1. If the frontend-design skill is already active in this claude.ai session, read its content as a base template, then customize it using Phase 1 call 5 answers before writing to docs/product/frontend-design.md. 2. If a local copy exists on the machine (common paths: ~/.agents/skills/frontend-design/SKILL.md, C:\Users\<user>\.agents\skills\frontend-design\SKILL.md, /mnt/skills/public/frontend-design/SKILL.md), read it as a base template and apply the same call 5 customizations. 3. If neither source is reachable, generate docs/product/frontend-design.md directly from the call 5 answers — do not use a generic minimal fallback. - Q11 (component library) → "Component System" section: library name, install command, import conventions, and Tailwind config notes (if applicable) - Q12 (design aesthetic) → "Visual Language" section: color tone, spacing density, border-radius scale, font personality, shadow depth - Q13 (layout pattern) → "Layout Patterns" section: primary nav structure, page skeleton template, responsive breakpoint strategy - Q14 (visual references / brand anchors) → "Reference Anchors" section: existing brand/UI to preserve, cited inspirations, and explicit "avoid" cues - Q15 (content density) + Q16 (theme preference) → "Preview Rendering Rules" section: density target, light/dark expectation, CTA hierarchy, and preview emphasis Log which strategy was used as a note in docs/learnings.md. For every frontend project, keep the UI artifact chain consistent: docs/product/frontend-design.mddocs/product/design.mddocs/product/design-preview.html. docs/product/frontend-design.md defines the global design system and style direction. docs/product/design.md translates that into the product-specific wireframe: overall app shell, navigation, global layout regions, and per-page/screen contracts. The HTML preview is a mid-fi styled static preview, not a pure wireframe and not production code.

  • GitBook / project-intro companion docs: for ALL new projects, generate docs/gitbook/ as part of Phase 3 scaffold — this is not conditional on user request. Treat it as a required parallel deliverable alongside AGENTS.md and PLAN.md. Generate and maintain a docs/gitbook/ markdown set by default (or the repo's existing docs root if one already exists). Minimum scope: landing page, product overview, problem/users, architecture/capabilities, quickstart, roadmap, and SUMMARY.md. Preferred starter shape: docs/gitbook/README.md, docs/gitbook/product-overview.md, docs/gitbook/target-users.md, docs/gitbook/architecture.md, docs/gitbook/quickstart.md, docs/gitbook/roadmap.md, and docs/gitbook/SUMMARY.md. Keep these pages derived from docs/PRD.md, ARCHITECTURE.md, docs/PLAN.md, and the current codebase state. When generating or revising docs/PLAN.md, add explicit GitBook tasks whenever a milestone changes product positioning, onboarding, architecture, integrations, or roadmap. For a new project, add an early baseline task to create the GitBook structure. For a retrofit, add a catch-up docs task or milestone if no coherent project-introduction docs exist yet.

Mode Selection

First, determine which mode to use. If the user uploads files or mentions an existing project, use Retrofit. If they describe something new, use Greenfield.

If unclear, ask with ask_user_input when structured prompt tools are available. If your runtime does not provide ask_user_input (for example Codex in a plain terminal flow), ask the same choice in prose and continue from the user's answer:

ask_user_input / prose equivalent:

  1. Mode (single_select): What's the starting point?

- Options: New project from scratch, Add harness to an existing project

If New project from scratch → read docs/agent/skill-greenfield.md and follow it If Add harness to existing project → read docs/agent/skill-retrofit.md and follow it

适合场景

01

用户想查找某类 Agent Skill 时

02

需要根据任务场景推荐可安装能力包时

03

需要对比不同来源的安装命令和来源信息时

能力概览

能力 1

按任务关键词查找相关 Skills

能力 2

展示可复制的安装命令

能力 3

保留来源站点、仓库和原始说明,方便继续核验

能力 4

展示第三方安全扫描或审计结果

安装后应在对应宿主中按原始 README 的触发条件使用;具体调用方式请以来源页面和 README 为准。

平台分布

Codex

37.51%
按下载量换算29

Claude

30.59%
按下载量换算24

Cursor

17.26%
按下载量换算13

Gemini CLI

9.76%
按下载量换算8

安全审计

Gen Agent Trust Hub

通过

Socket

可疑

Snyk

通过

权限和风险

执行命令

安装流程涉及命令执行,可能通过 npx skills add https://github.com/phlegonlabs/skills --skill harness-engineer-cli 联网下载 Skill 或依赖。用户安装前应确认命令来源、仓库内容和执行环境。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。来源安全扫描存在 warning/failed 结果,不能写成本站确认安全。当前只有一个来源,正式发布前建议补源仓库或其他目录站核验。

来源信息

继续浏览同类 Skills