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 skills add metahub-tech/agent-fleet --skill using-human-domgit clone --depth 1 https://github.com/metahub-tech/agent-fleetWrote 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/skills/metahub-tech/agent-fleet/using-human-dom)<a href="https://agentmods.dev/skills/metahub-tech/agent-fleet/using-human-dom"><img src="https://agentmods.dev/badge/skills/metahub-tech/agent-fleet/using-human-dom/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/metahub-tech/agent-fleet/using-human-dom"><img src="https://agentmods.dev/badge/skills/metahub-tech/agent-fleet/using-human-dom.svg" alt="Reviewed on agentmods" width="80" 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 5.1 | $0.00087 | $0.04796 |
| Opus 5.5 | $0.00035 | $0.01918 |
| Sonnet 5.5 | $0.00017 | $0.00959 |
| Haiku 4.5 | $0.00009 | $0.00480 |
Grade A, and why
using-human-dom 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 15d 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 — 174 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Using human_dom
human_dom is a DOM-perception companion to human_browser: it extends the same zero-trace, real-profile, OS-level operating model with a read-only DOM locator sourced from a Chrome extension content script.
The extension only reads — it never clicks, never mutates the DOM, never injects synthetic events. All input is still genuine OS-level: tap, type_text, press_key. The result is DOM-accuracy coordinates delivered as screen points in the same space as take_screenshot and tap.
三选一路由（routing）
| 场景 | 用哪个 |
|---|---|
| 真实账号/身份 + 需要 DOM 精确定位（文字/aria/placeholder 匹配） | human_dom（本 skill） |
真实账号/身份，DOM 拿不到（canvas、自定义控件、动态悬浮层）或 human_dom_locate 返回 suggest:"vision_locate" |
vision_locate（OCR 兜底），操作仍用 OS 级 tap/type_text |
| 自动化测试 / 抓取 / 隔离 profile / 无需真实身份 | agent_browser（Playwright，有 automation traces，独立 profile） |
核心判断：需要真实身份 → 选 human_dom 或 vision_locate 而非 agent_browser。需要 DOM 精度 → human_dom 优先；DOM 不可达 → OCR 兜底。
human_browser 上元素定位的优先级顺序
在 human_browser 打开页面后，定位元素按以下优先级尝试：
- human_dom（首选）：DOM 语义匹配，精确、0 token、抗动态遮罩，命中率高。
- vision_locate（OCR 兜底）：DOM 拿不到时（canvas/shadow DOM/动态悬浮层），OCR 识别屏幕文字返回坐标；操作仍用 OS 级 tap/type_text。
- 截图 + VLM 眼估（最后手段）：take_screenshot 后让模型目测坐标，仅当 OCR 也失败时才用。
| 方法 | 精度 | token 消耗 | 适用场景 |
|---|---|---|---|
| human_dom | DOM 像素精确 | 0 | 有文字/aria/placeholder 的标准元素 |
| vision_locate (OCR) | ~1px | 极少 | canvas、自定义控件、动态覆盖层 |
| 截图 + VLM | ~5–20px | 高（图片 token） | OCR 也失败的极端情况 |
实战分工：表单字段用 human_dom，自定义按钮常需 vision 兜底
real-machine（小红书 / 公众号发布）经验，钉成规范：
- 标准表单字段（标题
<input>、<textarea>、[contenteditable]正文）→ human_dom（human_dom_fill/human_dom_locate），DOM 精确、0 token。 - 页面自定义按钮 / 控件（底部固定栏的「发布」「暂存离开」、开关、话题下拉项、上传封面按钮等，多为
div[role=button]）→ 文字常不进 human_dom 文本索引（human_dom_locate返回suggest:"vision_locate"），改用vision_locate/vision_tap兜底，实测精准命中。 - 经验法则：输入走 human_dom；点按钮先 human_dom，拿不到立刻转 vision——不要在一个拿不到的按钮上反复
human_dom_locate。
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.
- 15d ago Changed · +6 lines 64d181f911d0
- 2mo ago First seen · 168 lines · 87 tokens per session scan A d9238002fc47
using-human-dom is a skill published in the GitHub repository metahub-tech/agent-fleet (3 stars, last pushed 16d ago), licensed Apache-2.0. It adds 87 tokens to every session and 4,796 once invoked, about $0.0003 per session on Opus 5.5. 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 2026-08-31.
using-human-dom compared with skills from other repositories
| Mod | Type | Stars | Last push | Licence | Scan | Tokens | Works with |
|---|---|---|---|---|---|---|---|
| using-human-dommetahub-tech/agent-fleet | Skill | 3 | 16d ago | Apache-2.0 | A | 87 | Claude Code, Codex |
| web-exfiltration-detectionTencent/AI-Infra-Guard | Skill | 7.4k | 2d ago | Apache-2.0 | A | 61 | Claude Code |
| browserwing-executorMemTensor/MemOS | Skill | 12k | 2d ago | Apache-2.0 | A | 42 | Claude Code, Codex |
| dev-browserMemTensor/MemOS | Skill | 12k | 2d ago | Apache-2.0 | A | 84 | Claude Code, Codex |
| web-searchopensquilla/opensquilla | Skill | 7.1k | 6d ago | Apache-2.0 | A | 76 | Claude Code, Codex |
| langbot-testinglangbot-app/LangBot | Skill | 19k | yesterday | Apache-2.0 | A | 58 | Claude Code, Codex |
| browser-automationPrism-Shadow/penguin-harness | Skill | 3.2k | yesterday | Apache-2.0 | A | 103 | Claude Code, Codex |