What is llm-answer-reference-compare?

A browser-based collector for saving complete answers and cited sources from DeepSeek, Doubao, Tencent Yuanbao, Tongyi Qianwen, and other listed Chinese AI services. It uses Playwright, a tool for controlling web browsers, and does not judge whether the answers are true.

Use it to collect AI answers, citation titles and URLs, source text, and browser evidence for comparisons or Fact-Check-X's answer-collection stage.

Comparing AI answers is difficult when citations, source links, or parts of an answer are missing. This keeps the original wording, citation markers, source order, and login or extraction failures visible for later fact-checking.

llm-answer-reference-compare — Skill for Claude Code and Codex by ASI2

Install

Getting it into your agent

It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.

Clone the repo

git clone --depth 1 https://github.com/ASI2030/Fact-Check-X

agentmods

npx agentmods add skills/asi2030/fact-check-x/llm-answer-reference-compare

Made for: Claude Code, Codex.

Or install fact-check-x, the plugin that ships this one along with the rest of its 5 skills.

Wrote 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.

README.md

[![agentmods](https://agentmods.dev/badge/skills/asi2030/fact-check-x/llm-answer-reference-compare/github.svg)](https://agentmods.dev/skills/asi2030/fact-check-x/llm-answer-reference-compare)

Your own site

<a href="https://agentmods.dev/skills/asi2030/fact-check-x/llm-answer-reference-compare"><img src="https://agentmods.dev/badge/skills/asi2030/fact-check-x/llm-answer-reference-compare/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.

Your own site · 80×15

<a href="https://agentmods.dev/skills/asi2030/fact-check-x/llm-answer-reference-compare"><img src="https://agentmods.dev/badge/skills/asi2030/fact-check-x/llm-answer-reference-compare.svg" alt="Reviewed on agentmods" width="80" height="20"></a>

Per session 88 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.

When invoked 1,564 The whole file, excluding the scripts and references it only reads on demand.

Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.

Origin original No closer match found in the catalogue.

Token cost

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	$0.00088	$0.01564
Opus 5	$0.00044	$0.00782
Sonnet 5	$0.00018	$0.00313
Haiku 4.5	$0.00009	$0.00156

Measured 9d ago against content hash 6cdb509ceaf0, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-09, from the pricing page.

Security

Grade A, and why

llm-answer-reference-compare 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 9d ago.

The scan reads SKILL.md. This mod also ships 29 executable files (assets/tool/dist/capture/auth-state.js, assets/tool/dist/capture/browser-session.js, assets/tool/dist/capture/generic-chat.js, …), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

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.

skills/llm-answer-reference-compare/SKILL.md · 106 lines

How it starts

The opening of the file, as written. The whole thing — 106 lines — stays where its author put it; the contents beside it link to each section on GitHub.

多平台回答与引用无损采集

本技能只负责采集，不拆知识点、不判断真假，也不调用任何模型 API。可信搜索仅用于按平台已经给出的来源标题与原始 URL 补全该来源全文，不用于另找证据、替换原始来源或提前判断真假；语义工作由当前承载技能的智能体完成。

输出原则

完整保存平台原答案，不摘要、不改写、不截断。
按平台原顺序保存引用标题、原始 URL、标记、摘要与已捕获正文。
正文有引用标记时保留标记与引用顺序；没有标记时明确记录为无显式标记。
原始 URL 是存证主键。normalizedUrl 只用于去重，禁止用搜索结果替换原始 URL 或正文。
页面只显示来源名称但不暴露 URL 时，先尝试展开来源控件获取真实链接；仍无链接则写入 sourceMentions，明确标记为“无 URL 来源标签”，不得伪造成参考文献。
平台同时提供逐句脚标和“参考 N 篇资料”等全局来源列表时，两类都必须采集并用 citationScope 区分；同一 URL 合并时保留更完整的标题与摘录。
页面声明全局来源数量时，必须展开并达到该数量。未抓全视为采集失败，进入自动重采和 Computer Use 接管，禁止以残缺引用继续知识点对比。
深知来源正文不足时，若安全环境已配置 TRUSTED_SEARCH_KEY，调用可信搜索并使用 return_full_content=true、simplified=false 补全与当前来源标题或 URL 匹配的同一材料全文；返回全文或段落直接用于判断，不再访问源网址二次抓取正文。可信搜索的 源网址 作为官方来源主链接，深知收录页作为辅助链接；不得把其他材料伪造成平台引用。
可信搜索未配置或未命中时，才直接读取来源页或提取 PDF；已绑定 PDF 仍拿不到可核验正文时必须失败关闭并进入正文提取、OCR 或 Computer Use 接管，不能把采集器的缺口判成平台幻觉。
登录、验证码、超时、页面结构变化和抽取失败必须成为显式状态，不能伪装成成功空回答。

准备运行时

cd assets/tool
npm ci --omit=dev
npx playwright install chromium
cd ../..

查看支持的平台：

node assets/tool/dist/cli.js platforms

首次使用某个平台时，先明确告诉用户“将打开浏览器，请完成登录，登录后再获取信息”。随后运行登录准备；程序检测到可提问界面后自动保存会话：

node assets/tool/dist/cli.js login --platform doubao

不得代替用户输入账号、密码、短信验证码或处理人机验证。登录会话保存在 ~/.fact-check-x/browser-profiles，不得进入技能包或核验报告；已有持久化登录会话后可直接采集。

豆包等未登录页面即使显示输入框，也不得视为登录完成。必须等待页面登录入口消失并确认可提问界面后才能提交问题；登录入口仍可见时只等待用户登录，不得预先填入问题。

采集一个或更多平台。默认等待回答确认生成完成，并在失败后自动重采 2 次：

node assets/tool/dist/cli.js run \
  --question "<用户原始问题>" \
  --platform dknowc-chat \
  --platform doubao \
  --out <run目录> \
  --headed \
  --interactive \
  --timeout 180000 \
  --retries 2

平台参数可以重复任意 N 次（N≥1）。内置 deepseek；深知晓地址可用 dknowc-chat=https://目标地址/ 覆盖，其他平台同理。例如：

node assets/tool/dist/cli.js run \
  --question "<用户原始问题>" \
  --platform dknowc-chat=https://测试5地址/ \
  --platform deepseek \
  --platform doubao \
  --out <run目录> \
  --headed --interactive

内置平台 dknowc-deep-research 对应 https://poc1.dknowc.cn/wlcb/shenzhimini-test5/。它不是普通单段采集：采集器先提交原问题并等待普通回答完整生成，再点击回答下方“深度研究”入口，接管新打开的可信溯源报告页，等待深度研究结果完整生成后提取答案、来源、截图和页面存证。它与 dknowc-chat 是两个独立平台结果；只有 dknowc-chat 具备后续可信锚点免查资格。

Read the full file on GitHub · 106 lines

Files

What ships with it

36 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

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.

9d ago First seen · 106 lines · 88 tokens per session scan A 6cdb509ceaf0

Subscribe to this mod's changes

llm-answer-reference-compare is a skill published in the GitHub repository ASI2030/Fact-Check-X (1 stars, last pushed 8d ago), licensed Apache-2.0. It adds 88 tokens to every session and 1,564 once invoked, about $0.0004 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 2026-08-31.

Browser automation Claude Code Research agent-skills ai-agents codex evidence fact-checking

web-scraping

Authorized web scraping with fallback cascades and access-failure handling. Use for social media, yt-dlp, CAPTCHA or 403 blocks.

jamditis/claude-skills-journalism · 33 tokens

video-download

Collects videos from public social accounts. Use to download or scrape from Twitter, TikTok, YouTube, Instagram, or Facebook.

jamditis/claude-skills-journalism · 30 tokens

browser-qa

Drive a real headless browser against a locally-served web UI and capture screenshot evidence, reporting step-level PASS/FAIL/BLOCKED with extracted DOM text and console errors. Use when a phase on the docker QA image must exercise a rendered UI and attach screenshots — it complements the text-evidence verify/qa-test…

nearform/lastlight · 69 tokens

demo

Record a short demo VIDEO of a PR or feature — drive the repo's web UI in a real headless browser, capture the session, and composite a titled, size-capped mp4 with ffmpeg. Use on the docker QA image when the deliverable is a playable demo clip (single walkthrough or before/after comparison), not a text/screenshot…

nearform/lastlight · 75 tokens

acquisition-graduation

Convert repeated dev-browser acquisition successes into durable Spotlight acquisition guidance. Use after a non-trivial browser acquisition works repeatedly and should become a reusable source/domain path without storing secrets, cookies, or fragile session details.

buriedsignals/spotlight · 47 tokens

touch-browser

Use for evidence-first public web research and claim verification with touch-browser. Trigger when the task needs browser-backed discovery, page-grounded citations, claim extraction, or safe unresolved outcomes instead of guessed answers. Do not use for private account automation, uncontrolled browsing, or final truth…

nangman-infra/touch-browser · 63 tokens