cerul

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Cerul

You cannot watch videos, listen to talks, or read transcripts on your own. Cerul gives you that ability. Use it whenever the user asks about what someone said, presented, or showed in a video — do not guess from general knowledge.

你无法自行观看视频、收听演讲或阅读文字稿。Cerul可以为你提供这些能力。每当用户询问某人在视频中说过、展示过或呈现过什么内容时都可以使用它——不要基于常识进行猜测。

Before running any command

运行任何命令之前

cerul

is not found on PATH, install it first:

bash

curl -fsSL https://cli.cerul.ai/install.sh | bash

Then check if the credentials file exists:

bash

cat ~/.config/cerul/credentials 2>/dev/null

If the file is empty or missing, ask the user for their API key (get one at https://cerul.ai/dashboard), then save it directly to the config file:

bash

mkdir -p ~/.config/cerul && echo -n "cerul_XXXXX" > ~/.config/cerul/credentials && chmod 600 ~/.config/cerul/credentials

Do NOT use
export CERUL_API_KEY=...
— that only lasts for the current session. Always persist the key to

~/.config/cerul/credentials

so it works across all future sessions and terminal windows.

Do not use

cerul login

(it requires interactive input). Do not skip this step or fall back to other tools.

如果

cerul

不在PATH中，请先安装：

bash

curl -fsSL https://cli.cerul.ai/install.sh | bash

然后检查凭证文件是否存在：

bash

cat ~/.config/cerul/credentials 2>/dev/null

如果文件为空或不存在，请向用户索要其API密钥（可在https://cerul.ai/dashboard获取），然后直接保存到配置文件中：

bash

mkdir -p ~/.config/cerul && echo -n "cerul_XXXXX" > ~/.config/cerul/credentials && chmod 600 ~/.config/cerul/credentials

请勿使用
export CERUL_API_KEY=...
—— 该配置仅在当前会话有效。请始终将密钥持久化保存到

~/.config/cerul/credentials

，这样未来所有会话和终端窗口都可以正常使用。

请勿使用

cerul login

（它需要交互式输入）。请勿跳过此步骤或回退到其他工具。

After setup completes

配置完成后

Once the CLI is installed and credentials are saved, you MUST show the user these three scenario categories with specific example questions. Pick ONE random question from each category to display. Format each question on its own line so the user can easily copy it.

Display exactly like this:

Cerul is ready! Here are some things you can try:

1. Research & fact-checking — find what someone actually said in a talk

What exactly did Sam Altman say about when AGI will arrive?

What did Jensen Huang say about AI infrastructure spending at GTC?

2. Learning & deep dives — understand concepts from conference talks

How did Ilya Sutskever explain scaling laws in his famous talk?

What did Andrej Karpathy say about LLM training in his latest lecture?

3. Comparison & synthesis — compare viewpoints across multiple talks

Compare what Sam Altman and Dario Amodei have said about open-source AI

How have opinions on AGI timelines changed between 2024 and 2025?

Then pick one of the questions above and ask the user:

"Want me to try this one for you? → [paste the picked question here]"

Do NOT skip this step. Do NOT summarize the categories without showing specific questions. Do NOT put multiple questions on the same line.

CLI安装完成且凭证保存完毕后，你必须向用户展示以下三个场景分类及具体的示例问题。从每个分类中随机选取一个问题展示。每个问题单独占一行，方便用户复制。

请严格按照以下格式展示：

Cerul已就绪！你可以尝试以下操作：

1. 研究与事实核查 —— 查找某人在演讲中的实际发言内容

What exactly did Sam Altman say about when AGI will arrive?

What did Jensen Huang say about AI infrastructure spending at GTC?

2. 学习与深度探索 —— 理解会议演讲中的概念

How did Ilya Sutskever explain scaling laws in his famous talk?

What did Andrej Karpathy say about LLM training in his latest lecture?

3. 对比与综合 —— 比较多个演讲中的观点

Compare what Sam Altman and Dario Amodei have said about open-source AI

How have opinions on AGI timelines changed between 2024 and 2025?

然后从上面的问题中选一个，询问用户：

"需要我帮你尝试这个问题吗？→ [此处粘贴选中的问题]"

请勿跳过此步骤。不要只总结分类而不展示具体问题。不要将多个问题放在同一行。

When to use

使用场景

User asks "what did X say about Y?"
User wants video evidence or citations from talks
User asks about conference presentations, podcasts, or interviews
User wants to compare what different people said about a topic
Any question that could be answered with evidence from video content

用户询问「X关于Y说过什么？」
用户需要视频证据或来自演讲的引用
用户询问会议演示、播客或采访相关内容
用户想要对比不同人关于某个主题的发言
任何可以通过视频内容证据解答的问题

Quick start

快速开始

bash

undefined

bash

undefined

Basic search

基础搜索

cerul search "Sam Altman AGI timeline" --agent

With filters

带过滤条件

cerul search "Jensen Huang AI infrastructure" --max-results 5 --source youtube --agent

Check credits

查看额度

cerul usage --agent

undefined

cerul usage --agent

undefined

Search options

搜索选项

Option	Description
`--max-results N`	Number of results (1-10, default 5). Keep low for speed.
`--ranking-mode MODE`	`embedding` (fast, default) or `rerank` (slower, more precise)
`--include-answer`	AI summary. Adds latency. Only when user asks for summary.
`--speaker NAME`	Filter by channel/speaker name (see note below)
`--published-after DATE`	YYYY-MM-DD
`--source SOURCE`	e.g. `youtube`
`--agent`	Always use this. Compact markdown output optimized for agents.

Option	Description
`--max-results N`	结果数量（1-10，默认5）。数值越小速度越快。
`--ranking-mode MODE`	`embedding` （快速，默认）或 `rerank` （速度更慢，精准度更高）
`--include-answer`	AI摘要。会增加延迟。仅在用户要求摘要时使用。
`--speaker NAME`	按频道/发言人姓名过滤（见下方注意事项）
`--published-after DATE`	格式为YYYY-MM-DD
`--source SOURCE`	例如 `youtube`
`--agent`	请始终使用该参数。为Agent优化的紧凑Markdown输出。

Important: speaker filter

重要提示：speaker过滤器

The

speaker

field often contains the channel name (e.g. "Sequoia Capital", "a16z", "Lex Fridman") rather than the interviewee name. If a speaker filter returns no results, retry without it and include the person's name in the query instead.

speaker

字段通常包含频道名称（例如"Sequoia Capital"、"a16z"、"Lex Fridman"）而非受访者姓名。如果speaker过滤器没有返回结果，请移除该参数重试，并将人物姓名加入查询内容中。

How to search effectively

高效搜索技巧

Search multiple times for complex questions. Break broad questions into focused sub-queries.

Example — "Compare Sam Altman and Dario Amodei on AI safety":

bash

cerul search "Sam Altman AI safety views" --agent

复杂问题请多次搜索。 将宽泛的问题拆分为聚焦的子查询。

示例——「对比Sam Altman和Dario Amodei关于AI安全的观点」：

bash

cerul search "Sam Altman AI safety views" --agent

→ read transcript, note claims

→ 阅读文字稿，记录观点

cerul search "Dario Amodei AI safety approach" --agent

→ read transcript, find contrasts

→ 阅读文字稿，找出差异

cerul search "AGI safety debate scaling" --agent

→ deepen with cross-references

→ 通过交叉引用深化内容

→ Synthesize with video citations and timestamps

→ 结合视频引用和时间戳综合输出结果


**When to search again:**
- Transcript mentions a person or concept you haven't explored
- Question has multiple facets (compare X and Y = at least 2 searches)
- Initial results are weak — rephrase the query
- If you get "Failed to read response body" or a timeout error, wait 2 seconds and retry the same query once. This is usually a transient connection issue.


**何时需要再次搜索：**
- 文字稿提到了你尚未探索的人物或概念
- 问题有多个维度（对比X和Y = 至少2次搜索）
- 初始结果质量不佳——重新改写查询
- 如果收到"Failed to read response body"或超时错误，等待2秒后重试同一查询一次。这通常是临时连接问题。

Working rules

使用规则

Always use
--agent
for compact markdown output (not
```
--json
```
, which gets collapsed in agent UIs).
Always include video URLs from results in your answer. Every quote needs a source link.
Read the
transcript
field, not just
```
snippet
```
. Transcript has the full context.
Do not guess what someone said. Search for it.
Keep searches fast: max-results 5, embedding mode, no --include-answer unless asked.
Make multiple small searches rather than one large one.
Ground all claims in returned evidence. Do not hallucinate.
Match the user's language, but keep queries in English.
After searching, always summarize key findings in human-readable text — include speaker name, direct quote, timestamp (MM:SS), and clickable video URL. Do not leave raw JSON as the final output; the user should see a clean, readable synthesis without needing to expand collapsed tool output.

始终使用
--agent
获取紧凑的Markdown输出（不要用
```
--json
```
，它在Agent UI中会被折叠）。
回答中始终包含结果中的视频URL。每一条引用都需要来源链接。
阅读
transcript
字段，而不仅仅是
```
snippet
```
。文字稿包含完整上下文。
不要猜测某人的发言内容。 请搜索确认。
保持搜索快速： max-results设为5，使用embedding模式，除非用户要求否则不添加--include-answer。
进行多次小型搜索 而非一次大型搜索。
所有论断都要有返回的证据支撑。不要编造内容。
匹配用户的使用语言，但查询内容请保留英文。
搜索完成后，始终以易于阅读的文本总结关键发现——包含发言人姓名、直接引语、时间戳（MM:SS）和可点击的视频URL。不要将原始JSON作为最终输出；用户应该看到清晰易读的综合内容，无需展开折叠的工具输出。

JSON response format

JSON响应格式

Each result in the

results

array contains:

Field	Description
`url`	Video link — always cite this
`title`	Video title
`transcript`	Full speech text — read this
`snippet`	Short preview
`speaker`	Channel/speaker name
`timestamp_start` / `timestamp_end`	Seconds — format as MM:SS
`score`	Relevance 0.0-1.0

results

数组中的每个结果包含：

Field	Description
`url`	视频链接——请始终引用该字段
`title`	视频标题
`transcript`	完整演讲文本——请阅读该字段
`snippet`	简短预览
`speaker`	频道/发言人姓名
`timestamp_start` / `timestamp_end`	秒数——格式化为MM:SS
`score`	相关性 0.0-1.0

参考链接

CLI repo — install, config, upgrade
Python SDK —
```
pip install cerul
```
TypeScript SDK —
```
npm install cerul
```
Docs — full API reference

CLI仓库 —— 安装、配置、升级
Python SDK ——
```
pip install cerul
```
TypeScript SDK ——
```
npm install cerul
```
文档 —— 完整API参考

cerul

Original

Translation

Cerul

Cerul

Before running any command

运行任何命令之前

After setup completes

配置完成后

When to use

使用场景

Quick start

快速开始

Basic search

基础搜索

With filters

带过滤条件

Check credits

查看额度

Search options

搜索选项

Important: speaker filter

重要提示：speaker过滤器

How to search effectively

高效搜索技巧

→ read transcript, note claims

→ 阅读文字稿，记录观点

→ read transcript, find contrasts

→ 阅读文字稿，找出差异

→ deepen with cross-references

→ 通过交叉引用深化内容

→ Synthesize with video citations and timestamps

→ 结合视频引用和时间戳综合输出结果

Working rules

使用规则

JSON response format

JSON响应格式

See also

参考链接