Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/episodeyu/harnesssmith/claude-mdgit clone --depth 1 https://github.com/EpisodeYu/HarnessSmithWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/instructions/episodeyu/harnesssmith/claude-md)<a href="https://agentmods.dev/instructions/episodeyu/harnesssmith/claude-md"><img src="https://agentmods.dev/badge/instructions/episodeyu/harnesssmith/claude-md.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.02811 | $0.02811 |
| Opus 5 | $0.01406 | $0.01406 |
| Sonnet 5 | $0.00562 | $0.00562 |
| Haiku 4.5 | $0.00281 | $0.00281 |
Grade A, and why
HarnessSmith CLAUDE.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 128 lines — stays where its author put it; the contents beside it link to each section on GitHub.
CLAUDE.md — HarnessSmith 项目守则
Agent 进入本项目(vibe coding 模式)入场必读。本文件是 Agent 工作的硬约束。 项目定位 / 范围 / 决策 / 开发节奏与切片门禁见
docs/02-development/00-overview.md。Tradeoff:偏向"先把事情做对"。显然小到不值得讨论的操作自己判断,别让流程把简单任务搞复杂。
0. 协作模式与角色边界
本项目主要由 Agent 开发,人只负责决策、验证、以及"只能人做"的环境配置。
| 角色 | 负责 | 不做 |
|---|---|---|
| Agent(你) | 写生成器与模板代码、写/跑测试、调试、维护文档与变更记录、按 slice 自驱推进、测试全绿后直接 commit + push main、写完成报告 |
不替人拍板产品决策、不动用钱/联网批量/破坏定位的事、不 --force push、不在测试未绿时推 main |
| 人 | 方向与验收、slice 关键决策点签字、环境与密钥配置(见 §0.1)、对外发布 | 不写代码、不做 Agent 已能自动化的事 |
一句话:Agent 负责"怎么做",人负责"做不做、装环境、给 key、何时发布"。 开发文档里写"开发者要做 X"一律理解为"你(Agent)要做 X,由人按切片节奏验收"。
0.1 只能由人配置的环境(Agent 不代劳)
- 安装本机工具链:
uv、docker(Agent 不改系统级配置、不装系统包)。 - API key /
.env真实值:Agent 只维护.env.example,绝不写入真实.env、绝不把 key 写进任何被 git 跟踪的文件。日常开发用 mock LLM,不需要真 key;只有"真实 LLM 冒烟 / 对外演示"才需要人提供 key。 - PyPI 发布凭证、GitHub 仓库设置、对外发布。
Agent 需要某个 key / 环境时:在 plan 阶段列出所需
.envkey 名与用途,停下来请人配置,不要自己编造或跳过。
1. Think Before Coding
不要假设。不要藏起困惑。把权衡显式说出来。 动手前:声明假设,不确定就问;多种合理解读时列出来让人选;有更简单的方案先说一声。宁可开工前多问 1 个问题,不要跑了 30 分钟才发现方向错了。
2. Thin First(薄优先 —— 本项目核心卖点,硬约束)
- 默认产物模板保持极薄:核心循环目标 150–300 行,整体远小于一个框架。
- 不做没要求的功能;一次性代码不先抽象;不为"以后可能要"加灵活性。
- 高级能力(RAG / MCP / context 策略 / Web)只通过 spec 开关生成,不塞进默认产物。
- 200 行能压到 50 行就重写。
3. Surgical Changes
只动该动的,只清理自己制造的烂摊子。不"顺手优化"无关代码或格式;不重构没坏的东西;沿用现有风格;发现 dead code 提一句,别擅自删。
4. 两层心智:生成器 vs 生成产物(本项目特有)
你写的是一个生成器,它渲染出独立的生成产物仓库。任何时候分清你在改哪一层:
- 改
harnessmith/(生成器本体):spec / 渲染引擎 / CLI / 向导 / catalog / presets。 - 改
harnessmith/templates/(产物模板):渲染后才是用户拿到的代码。 - 测试必须覆盖"生成产物"本身:生成 →
uv sync→pytest→ mock LLM 跑通一次工具调用,而不只是测生成器。
5. 目标驱动 + 测试硬门槛(按生成器项目定制)
"完成"的硬门槛见 §5.1,没达到 = 没完成,别用"逻辑简单不用测"或"先合后补"做借口。 例外:纯文档 / 纯注释 / 纯重命名(改名工具已覆盖全部引用)可不加测试,但要在完成报告里点名说明。
5.1 黄金测试是"完成"的硬门槛
任何被宣称完成的功能必须:
- 新增 / 改动的生成器或模板代码有自动化测试(unit 或 integration)。
- 黄金路径绿:用示例 / preset spec 生成项目 →
uv sync && pytest全绿 → mock LLM 跑通一次 function-calling(含一次工具调用)。 - 断言生成的
pyproject.toml不含langchain/langgraph/adk。 - 可运行性自检绿:生成后冒烟(
uv sync+ import + mock 跑一步 +pytest -q)。 ReadLints(IDE 诊断)无新增 error / warning。
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago First seen · 128 lines · 2,811 tokens per session scan A 55d5e10f0019
HarnessSmith CLAUDE.md is an instructions file published in the GitHub repository EpisodeYu/HarnessSmith (22 stars, last pushed 1mo ago), licensed MIT. It adds 2,811 tokens to every session, about $0.0141 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
Sponsio CLAUDE.md
Claude Code instructions for SponsioLabs/Sponsio, covering agent guide for sponsio, what sponsio is, positioning, what to read first and repository map.
holaOS CLAUDE.md
Claude Code instructions for holaboss-ai/holaOS, covering holaos — repo guide and hola — debug the hola (pi) brain from the cli.
nulnul-harness AGENTS.md
AGENTS.md instructions for SeoNaRu/nulnul-harness: This repository builds plugins/nulnul-harness/, a Codex plugin that selects the strongest justified project-local capability system for the requested outcome, verifies the user's work, removes non-contributing setup, and evolves from measured outcomes without…
pixir CLAUDE.md
Instructions for Ranvier-Technologies/pixir, covering pixir harness - legacy agent guide, current map and commands.
harmont-cli CLAUDE.md
Instructions for harmont-dev/harmont-cli, covering testing, documentation, dsl and keep the sdk, hm init templates, and docs in sync.
browser-agent CLAUDE.md
Instructions for uiuing/browser-agent, a project described as: An open-source AI agent for the browser. Bring any model. It plans in the side panel, acts on real pages through tools, and verifies every claimed result against the live DOM — with guardrails and reusable skills.