hertzflow

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

hertzflow — Crypto trade-decision intelligence (single skill, multi-domain)

hertzflow — 加密货币交易决策智能工具（单技能，多领域）

hertzflow

is HertzFlow's umbrella skill for on-chain + CEX trading research. Like Surf, it's one skill with internal routing — read the table below to see which sub-domain handles the user's request, then follow that sub-domain's

INSTRUCTIONS.md

hertzflow

是HertzFlow旗下用于链上+CEX交易研究的总括性技能。与Surf类似，它是一个内置路由的单技能——请阅读下方表格确定哪个子领域处理用户请求，然后遵循该子领域的

INSTRUCTIONS.md

文档。

🌐 Language — MANDATORY, read first

🌐 语言设置 — 强制要求，请首先阅读

Detect the language of the user's most recent message and use that language for everything: your conversation replies AND the generated report. This is not optional and not a preference the user has to ask for.

Message contains CJK characters (
```
一-鿿
```
, kana, hangul) → operate in Chinese (
```
zh
```
): reply in Chinese, run the pipeline with
```
--lang zh
```
.
Otherwise (English / Latin-script prompt) → operate in English (
```
en
```
): reply in English, run the pipeline with
```
--lang en
```
.

Supported report languages are

zh

and

en

(static, human-curated packs). For a prompt in any other language, reply in that language but generate the report in

en

(closest supported), and say so in one line.

The report-body language is locked at pipeline time via

forensic_pipeline.py --lang {zh|en}

— it cannot be changed at render time, and narrative fills must be authored in the same language. See the "Report language" section of

alpha/INSTRUCTIONS.md

for the exact rule. Getting this wrong forces a full pipeline re-run (re-spends Surf credits), so set it from the user's language up front.

检测用户最新消息的语言，并全程使用该语言：包括对话回复和生成的报告。 这是强制要求，无需用户主动提出偏好。

消息包含CJK字符（
```
一-鿿
```
、假名、谚文）→ 使用中文（
```
zh
```
）：用中文回复，通过
```
--lang zh
```
参数运行流水线。
其他情况（英文/拉丁字母输入）→ 使用英文（
```
en
```
）：用英文回复，通过
```
--lang en
```
参数运行流水线。

支持的报告语言为

zh

和

en

（静态人工整理包）。若输入为其他语言，用该语言回复，但生成英文报告（最接近的支持语言），并单独用一行说明此情况。

报告正文语言在流水线运行时通过

forensic_pipeline.py --lang {zh|en}

锁定——无法在渲染阶段更改，叙事内容填充必须使用相同语言。具体规则请参见

alpha/INSTRUCTIONS.md

中的“报告语言”章节。若设置错误，将需要重新运行整个流水线（消耗Surf积分），请从用户输入语言出发提前设置正确参数。

Routing table

路由表

If the user input matches…	Sub-domain	Instructions file
Raw 0x-prefixed 40-hex EVM CA ( `0xea37a8de…` )	`alpha` (Binance Alpha forensic)	`alpha/INSTRUCTIONS.md`
Solana base58 CA (e.g. `Grass7B4R…` )	`alpha` (HOLDER_SNAPSHOT mode)	`alpha/INSTRUCTIONS.md`
Ticker of a token currently on Binance Alpha ( `$JCT` , `$VELVET` , etc.)	`alpha` — first resolve ticker → CA via `surf search-token --q <ticker>` , then dispatch	`alpha/INSTRUCTIONS.md`
Natural-language phrasing about Alpha distribution / 内幕出货 / 链上派发 / quiet insider / chip structure	`alpha`	`alpha/INSTRUCTIONS.md`
Perp symbol (e.g. `BTCUSDT` ) + analysis intent	`perp` (not yet shipped)	— coming soon
Bridge protocol name + audit intent	`bridge` (not yet shipped)	— coming soon
HertzFlow's own contract address + diagnostic intent	`core` (not yet shipped)	— coming soon

用户输入匹配…	子领域	说明文档
原始0x前缀的40位十六进制EVM合约地址（CA，如 `0xea37a8de…` ）	`alpha` （Binance Alpha链上取证）	`alpha/INSTRUCTIONS.md`
Solana的base58格式合约地址（如 `Grass7B4R…` ）	`alpha` （HOLDER_SNAPSHOT模式）	`alpha/INSTRUCTIONS.md`
当前Binance Alpha上的代币代码（如 `$JCT` 、 `$VELVET` 等）	`alpha` — 先通过 `surf search-token --q <ticker>` 将代码解析为合约地址（CA），再分发请求	`alpha/INSTRUCTIONS.md`
涉及Alpha筹码派发/内幕出货/链上派发/隐秘内幕方/筹码结构的自然语言表述	`alpha`	`alpha/INSTRUCTIONS.md`
永续合约交易对符号（如 `BTCUSDT` ）+ 分析需求	`perp` （尚未上线）	— 即将推出
跨链桥协议名称 + 审计需求	`bridge` （尚未上线）	— 即将推出
HertzFlow自身合约地址 + 诊断需求	`core` （尚未上线）	— 即将推出

First action when invoked

触发后的首个操作

Identify the sub-domain from the routing table above (default:
```
alpha
```
).
```
cat <sub-domain>/INSTRUCTIONS.md
```
(paths are relative to this skill's install root — at runtime that's
```
~/.claude/skills/hertzflow/
```
or the equivalent agent skill dir, NOT the repo root. So for the alpha sub-domain, read
```
alpha/INSTRUCTIONS.md
```
.) Or read it via the agent's file-read tool) — that file is the full per-domain SKILL.md, including onboarding (Surf CLI install + API key prompt), prerequisites, 3-step workflow (
```
forensic_pipeline.py
```
→ LLM fill →
```
render_report.py
```
), writable-slot guide, validator rules, output schema.
Follow the instructions in that file verbatim. Do not re-derive the workflow from this router file — the per-domain instructions are authoritative and version-locked.

根据上述路由表确定子领域（默认：
```
alpha
```
）。
读取
```
<sub-domain>/INSTRUCTIONS.md
```
文件（路径相对于该技能的安装根目录——运行时为
```
~/.claude/skills/hertzflow/
```
或对应的Agent技能目录，而非代码仓库根目录。例如alpha子领域，读取
```
alpha/INSTRUCTIONS.md
```
）。可通过Agent的文件读取工具完成——该文件是各子领域完整的SKILL.md文档，包含入门指导（Surf CLI安装+API密钥提示）、前置条件、三步工作流（
```
forensic_pipeline.py
```
→ LLM填充 →
```
render_report.py
```
）、可编辑模块指南、验证规则、输出 schema。
严格遵循该文件中的说明操作。请勿从此路由文件推导工作流——各子领域的说明文档是权威且版本锁定的。

Why this skill is structured this way

该技能的架构设计原因

Single slash
/hertzflow
: HertzFlow is one product brand covering multiple research domains (Alpha forensic, perp metrics, bridge audit, HertzFlow core). Users invoke one slash, the router handles dispatch.
Description bait is wide: the description (above) tells the agent to auto-trigger on natural-language mentions of insider distribution, chip structure, etc. — like Surf, users don't have to type the slash explicitly.
Sub-domain isolation: each sub-domain lives in its own folder with its own
```
INSTRUCTIONS.md
```
, helpers, and version tracking. Adding
```
perp/
```
later does not require touching
```
alpha/
```
code.

单一命令入口
/hertzflow
：HertzFlow是覆盖多个研究领域（Alpha链上取证、永续合约指标、跨链桥审计、HertzFlow核心分析）的单一产品品牌。用户调用一个命令，路由系统处理分发。
触发范围广泛：上方描述告知Agent，当用户提及内幕筹码派发、筹码结构等自然语言内容时自动触发——与Surf类似，用户无需明确输入命令即可触发。
子领域隔离：每个子领域拥有独立文件夹，包含专属
```
INSTRUCTIONS.md
```
文档、辅助工具和版本跟踪。后续添加
```
perp/
```
子领域无需改动
```
alpha/
```
代码。

Trust boundaries (apply to every sub-domain)

信任边界（适用于所有子领域）

Pipeline (Python). Deterministic. Owns all SQL, on-chain reads, evidence-graph IDs. The only layer that touches data.
LLM (any model). Fills narrative slots inside structured writable fields. Cannot write SQL. Cannot pick the verdict. Cannot decide thresholds.
Validator. Rejects LLM writes outside the writable slots, references to data not in the evidence graph, duplicated boilerplate, and any attempt to overwrite locked pipeline output.

Every sub-domain's

INSTRUCTIONS.md

must respect these three boundaries.

流水线（Python）：具有确定性。负责所有SQL编写、链上数据读取、证据图谱ID生成。是唯一接触数据的层级。
LLM（任意模型）：仅在结构化可编辑字段中填充叙事内容。无法编写SQL，无法判定结论，无法设置阈值。
验证器：拒绝LLM在可编辑模块外写入内容、引用证据图谱外的数据、重复模板内容，以及任何覆盖流水线锁定输出的尝试。

所有子领域的

INSTRUCTIONS.md

文档必须遵守这三个信任边界。

Upgrade

升级

bash

npx skills update hertzflow

That pulls the latest revision from

main

HertzFlow/hertzflow-skills

. Run after major releases (

v0.9.x

→

v1.0.0

etc.). Current version is in the frontmatter (

metadata.version

) and

alpha/v06/helpers/_version.py

bash

npx skills update hertzflow

该命令将从

HertzFlow/hertzflow-skills

仓库的

main

分支拉取最新版本。请在重大版本发布后运行（如

v0.9.x

→

v1.0.0

等）。当前版本信息位于前置元数据（

metadata.version

）和

alpha/v06/helpers/_version.py

文件中。