HertzFlow Skills — a trading toolkit curated and developed by the HertzFlow research team, designed to give traders faster on-chain + CEX insight and support better trade-decision calls. Actively maintained and expanding.

English · 中文

hertzflow-skills is the umbrella repository for HertzFlow's open-source AI-agent trading toolkit. Each skill inside this repo plugs into your AI coding agent (Claude Code, Cursor, Codex CLI, Windsurf, and ~50 others) and gives the agent a specific research capability — on-chain forensics, CEX intelligence, perp metrics, and more.

You install the whole suite once with a single command, and every sub-skill lives in its own folder under skills/. New capabilities are added over time without breaking what you've already installed.

More sub-skills from the HertzFlow research team are in active development and will be added to this same /hertzflow slash as they ship — the skill is designed so new capabilities slot in without you needing to reinstall anything.

If you trade on Binance Alpha, you've probably seen this pattern: a token launches, runs up, and then bleeds for weeks. The people who saw the bleed coming weren't reading tea leaves — they were reading the on-chain data. They saw the project push supply into ten wallets before the launch. They saw two of those wallets quietly drip-feed into the LP. They saw the other eight sitting there, ready to dump on any rally.

HertzFlow's Alpha sub-skill makes that view available with one command. Paste the contract address, wait 2–10 minutes (depending on token activity and surf cache state), and get back a trader-facing report: who currently holds what, how the supply was distributed, which wallets to watch from here, the token's current on-chain state (5-tier: clean / watch / operator-dumped / dormant-insider-risk / recent-distribution), and a 7-dimension at-a-glance TL;DR covering chip structure, insider sell-out, volume quality, supply risk, and market stage.

You do not need to know what a "mint event" or a "fan-out" is. You don't need to keep ten BscScan tabs open. You don't need to write SQL. The sub-skill does the on-chain forensic work in the background and hands you a plain-language report.

For any Binance Alpha–listed BSC token, the skill answers the six questions a serious trader actually has to ask:

1. Is the inner circle dumping right now? The report splits supply movement into three phases: pre-launch OTC seeding, downstream carrier hops, and the most recent 72 hours of large transfers. You see at a glance whether distribution is finished or still active — and where in that cycle the token is.

2. How much can you actually buy without moving the price? The skill pulls Binance Alpha's official depth API (not aggregator estimates) and computes the maximum single-tranche buy under 5 % slippage. Aggregator-based retail tools (1inch / DEX Screener etc.) only see DEX pool depth and miss Binance's internal MM book, so they often understate true tranche capacity by an order of magnitude or more; this report gives you a number you can size against.

3. Who currently holds what? Holders grouped by role: project deployer / project lockup (vesting + multisig + treasury + DEX-infra + CEX custody, Arkham confirmed) / quiet insider (never moved) / partial dumper / dumper-destination (retail receiving downstream) / DEX pool / other (uncategorized retail). Each row shows current balance, percent of supply, and dollar value. You instantly see how much "hidden" supply is still parked, waiting for a rally.

4. Where did the supply come from? An m6 lineage table (tucked into a <details> fold inside the holdings section) lists every wallet that received tokens from the project deployer — how much they got, how much they still hold, and what percentage they've already dumped. You can tell apart project vesting addresses, market makers, and wallets that look like retail but were funded directly by the project.

5. Which wallets should you watch from here? You get a 15–20 wallet monitoring list with role labels and alert conditions, exported as JSON that pastes directly into Binance Wallet or OKX address-tracking. The moment a quiet insider starts moving, you know.

6. What's this token's on-chain state right now? The report's Decision summary assigns the token to one of 5 chain-state tiers — clean / watch / operator-dumped / dormant-insider-risk / recent-distribution — plus a 0–10 risk score and a max single-tranche entry size pulled from Alpha's depth API. The state tier + 7-dimension at-a-glance TL;DR is what you actually trade on; the skill deliberately does not give buy / sell instructions — it describes the chain, you make the call.

The goal is simple: give a non-engineer trader the same on-chain read that a full-time chain analyst would produce. End-to-end runtime is typically 2–10 minutes depending on token activity.

Most Alpha tokens are on BSC, but Alpha also lists tokens on other EVM chains — the skill runs full on-chain forensics on all of them. Solana is the only exception, called out below.

Full historical forensic — pre-launch OTC seeding, downstream carrier hops, anomaly waves, wash-infrastructure detection, cross-sym operator registry — runs on all 6 surf-SQL-covered EVM chains. The pipeline auto-routes SQL to the right per-chain tables via chain_router.py; no manual chain flag needed.

Solana tokens run in HOLDER_SNAPSHOT mode: Alpha API realtime data

• top holders + Arkham labels still come through, but every SQL-driven detector section (m6 lineage, real-distribution sell-out quantification, wash detection, 72h anomaly, TGE LP-first-trade) is skipped, with a prominent banner at the top of the report.

Non-EVM, non-Solana tokens (Sui / TRON / etc.) return realtime price / vol / FDV only — the forensic sections are entirely skipped.

npx skills add HertzFlow/hertzflow-skills

That command always installs the latest stable revision from main. You do not need to pin a version. To upgrade later:

npx skills update hertzflow

The skill is delivered through the vercel-labs/skills universal installer.

1. Surf account. Register a free account at https://agents.asksurf.ai/, install the Surf CLI per the official docs, then:

   surf auth --api-key sk-...        # persistent
   # or
   export SURF_API_KEY=sk-...        # env var

2. curl and jq. Standard on macOS / Linux. On Windows, use WSL or Git Bash.

The free anonymous Surf tier cannot complete a full report — register for a free account to get the bonus credits needed.

CA + alpha_listing_date
        ↓
forensic_pipeline.py     (Python, ~1–8 min, deterministic — no LLM)
  emits skeleton.json     locked data + <LLM_NARRATIVE_PLACEHOLDER> slots
        ↓
LLM fill                 (your agent fills ~100-250 narrative slots)
  emits filled.json
        ↓
validate_report_data.py  (rejects out-of-bounds writes, hallucinations,
                          duplicated narrative, missing provenance)
        ↓
render_report.py         (Jinja2, deterministic)
  emits report.md + monitoring/*.json

Three strict trust boundaries:

• Pipeline (Python). Owns ~70 % of the report — all SQL, on-chain reads, holder distribution, evidence graph IDs. Only layer that touches data.
• LLM (any model). Fills ~100–250 narrative slots (varies by token activity and m6 lineage size). Cannot write SQL, cannot pick the verdict, cannot decide thresholds.
• Validator. Rejects any LLM write outside its slots, any reference to data not in the evidence graph, and any duplicated boilerplate.

Result: different LLMs running the same skeleton produce the same verdict. Cross-model convergence tests show ~91 % leaf-level identity and 100 % match on verdict and action enums.

Reports ship in two languages — Simplified Chinese (default) and English. Both forensic_pipeline.py and render_report.py accept a --lang {zh,en} flag, which must match between the two calls.

python3 skills/hertzflow/alpha/v06/forensic_pipeline.py \
    --ca 0x... --out-dir ./out/ --lang en

python3 skills/hertzflow/alpha/v06/render_report.py \
    --skeleton ./out/.work/skeleton.json \
    --filled   ./out/.work/filled.json \
    --out      ./out/report.md \
    --lang     en

out/
├── report.md                              # user-facing report (Markdown)
├── monitoring/
│   ├── monitoring_paste.json              # Binance / OKX paste-import
│   ├── monitoring_gmgn.txt                # GMGN paste (bare addresses)
│   ├── monitoring_gmgn_quoted.txt         # GMGN paste (quoted form)
│   ├── monitoring_binance_web3.csv        # Binance Web3 CSV upload
│   ├── monitoring_okx.csv                 # OKX CSV upload
│   └── monitoring_wallets_full.json       # analytics (full record set)
└── .work/                                 # hidden intermediates
    ├── skeleton.json
    └── filled.json

The report is a single Markdown file structured top-to-bottom. If you only look at one thing, look at 🎯 At-a-glance at the very top — it compresses the entire report into 7 lines. The table below maps to the actual section headings you'll see in the rendered report (ordered as they appear).

If the report header carries a ⚠️ N narrative quality warnings line, the structural data is fine but the narrative slots that produced those warnings should be re-filled by your LLM agent for a cleaner read. The data itself is locked and validated either way.

The CEX listing tier is the single most important upstream catalyst classifier in the report — it determines what kind of price action you should expect and which exits are available.

The tier is computed from real listing data in the report header, not from the listing time. Look for the tier field in the header line.

The monitoring/ folder is the actionable piece — the wallets the report tells you to watch, formatted for direct import into the four trackers most Alpha traders already use. Paste route is more reliable than file upload — try paste first.

1. Open the Binance Wallet app → Address tracking / Track address.
2. Choose paste / import as JSON (Binance UI calls this "GMGN format").
3. Open monitoring/monitoring_paste.json, select all, copy, paste into the dialog. Each row has {address, name, emoji}; the name carries the role label (e.g. H-Deployer-3a6dc).
4. Confirm — addresses appear in your watch list with role labels.

If paste fails, fall back to monitoring_binance_web3.csv via the file-upload route (the web wallet supports CSV, not JSON).

1. OKX Wallet → Watchlist / Track address.
2. Choose paste-import. The same monitoring_paste.json works.
3. If paste fails, upload monitoring_okx.csv instead (columns: Network, Address, Label, Note).

GMGN's bulk import widget is likely Solana-only today — paste attempts on BSC 0x… addresses have returned "Invalid format" in our testing. If GMGN bulk import works for BSC for you, try monitoring_gmgn.txt (bare list) first, then monitoring_gmgn_quoted.txt (quoted array) as a fallback. Otherwise add wallets to GMGN one-by-one or stick with Binance / OKX.

This is the full record set — every wallet the pipeline identified, with current balance, dumped %, role enum, evidence graph reference, and recent-72h activity flag. Use this if you want to script your own alerts or feed the data into a spreadsheet. It is not importable into any tracker UI.

Wallet label format is <SYM>-<ROLE>-<addr5> (e.g. H-Deployer-3a6dc, H-Dumper-1de14), ≤25 chars to clear known tracker character limits. Role tokens: Deployer / Dumper / PDumper / Quiet / LP / Anomaly / Other.

MIT — free to use, modify, and redistribute. Attribution appreciated but not required.

Built by HertzFlow and released to the community. HertzFlow is a permissionless perpetuals exchange on BNB Chain; this skill is part of the open-source toolkit we use internally and contribute back to the ecosystem.

hertzflow-skills 是 由 HertzFlow 研究团队整理开发的交易用户 工具集, 旨在帮助交易者更快地获得洞察, 辅助交易决断。持续更新 维护中。

每个 sub-skill 都嵌入你的 AI coding agent (Claude Code / Cursor / Codex CLI / Windsurf / Cline / Continue / Aider 等 50+ 种), 给 agent 加一项专属研究能力 — 链上取证、CEX 情报、永续盘口、跨链桥 审计等。

一条命令装整套, 各 sub-skill 在 skills/ 下独立子目录, 未来加新 能力不影响你已经装好的部分。

HertzFlow 研究团队还在持续开发新 sub-skill, ship 后会自动并入同一个 /hertzflow slash — 架构上设计成新能力直接 slot in, 老用户不需要 重装任何东西。

做 Alpha 交易的, 应该见过这个模式: 一个币上线, 拉一波, 然后阴跌 几周。提前知道要阴跌的人不是看 K 线看出来的 — 他们看的是链上 数据。他们看到项目方上线前把筹码推到 10 个地址、看到其中 2 个 悄悄往 LP 里 drip、看到剩下 8 个还在那等着, 等行情一拉就准备砸。

HertzFlow 的 Alpha sub-skill 让你也能看到这一层, 一条命令搞定。 把合约地址粘进你的 AI agent, 等 2-10 分钟 (看代币活跃度 + surf 缓存命中情况), 拿到一份给交易者看的报告: 现在谁手里有多少 筹码、筹码是怎么派发的、接下来要盯哪些钱包、当前链上状态属于 5 档里的哪一档 (无显著触发 / 观察中 / 派完离场 / 潜伏内幕未派 / 近期 派发), 以及一个 7 维度 一屏结论 TL;DR (筹码结构、内幕套现情况、 成交质量、供应风险、盘口阶段) — 报告只描述链上, 不给买卖建议, 操作判断由你结合自己仓位 + 风险偏好决定。

你不需要懂链上数据。不需要知道什么叫 "mint event"、"fan-out"。 不需要同时开十个 BscScan 盯。不需要写 SQL。Sub-skill 在后台跑完 链上取证, 给你一份大白话报告。

对任意一个 Binance Alpha 上线的 BSC 代币, skill 回答严肃交易者 必须问的 6 个问题:

1. 内幕现在是不是在出货? 报告把筹码动向分成三段: 上线前 OTC 铺货 / 下游 carrier 接力 / 最近 72 小时大额转账。一眼看出派发是已经结束, 还是在进行中, 以及代币处于这个周期的哪个位置。

2. 不滑点你到底能买多少? Skill 直接拉 Binance Alpha 官方深度 API (不是聚合器估算), 算出 5% 滑点下最大单笔买入金额。聚合器类散户工具 (1inch / DEX Screener 等) 只看 DEX 池深度, 看不到 Binance 内部 MM 的盘口, 经常低估真实可吃的 size 至少一个数量级。报告给你一个能拿来 size 的真实数字。

3. 当前谁手里拿着多少? 持仓按角色分组: 项目方部署钱包 / 项目方+基建+分发池 (vesting + 多签 + treasury + DEX 基建 + CEX 托管, Arkham 已确认) / 潜伏 内幕 (从未动过) / 分发中内幕 / 散户接收 (内幕下游) / DEX 主池 / 其他 (未分类散户)。每行有当前余额、占总供应比例、美元价值。 一眼看出还有多少"隐藏"筹码停在那, 等行情拉起来出。

4. 筹码是从哪来的? 一张 m6 内幕谱系表 (塞在持仓分布段的 <details> fold 里) 列出每个收过项目方分发的钱包: 收了多少、当前还剩多少、已派出 %。 你能区分项目方 vesting 地址 / 做市商 / 表面散户但 实际是项目方关联钱包。

5. 接下来应该盯哪些钱包? 你拿到一份 15-20 钱包的监控列表, 带角色标签 + 告警条件, 导出 为 JSON 直接粘进 Binance Wallet 或 OKX 地址追踪。沉默 的内幕一动手, 你立刻收到提醒。

6. 这个代币现在链上是什么状态? 报告的 决策摘要 段 给出 5 档链上状态判定 (无显著触发 / 观察中 / 派完离场 / 潜伏内幕未派 / 近期派发) + 风险评分 (0-10)

   • 进场上限 (LP 5% 深度). 状态档位 + 一屏结论 7 维度才是你真正 交易要看的; skill 故意不给"买 / 卖 / 持有"指令 — 报告只 描述链上, 你自己结合仓位 + 风险偏好做判断。

目标很简单: 让一个不懂链上的 Alpha 交易者, 拿到一份等同于全职链上 分析师产出的报告。完整流程 (pipeline + LLM 填充 + render) 通常 2-10 分钟跑完。

跟你打交道最多的是 BSC, 但 Alpha 也会上其他 EVM 链的币 — 这些 skill 都能跑全套链上取证。Solana 例外, 单独说明:

完整链上取证 — 上线前 OTC 铺货、下游 carrier 接力、异动波次、对敲 设置检测、跨币种庄家集合 — 在 6 条 surf SQL 覆盖的 EVM 链上全部 支持, pipeline 自动把 SQL 路由到对应链的数据表, 不需要手动指定。

Solana 代币走 HOLDER_SNAPSHOT 模式: Alpha API 实时数据 + 头部 持币人 + Arkham 标签仍能拿到, 但所有依赖 SQL 的检测段 (m6 谱系 / 真实派发卖出量化 / 对敲检测 / 72h 异动 / TGE LP 首笔) 跳过, 报告头 部会有明显提示。

非 EVM / 非 Solana 代币 (Sui / TRON 等) 只返回实时行情字段, 链上 取证全部跳过。

npx skills add HertzFlow/hertzflow-skills

这条命令永远拉 main 分支 HEAD = 最新 stable。不要带 @v0.X.X 后缀, 除非有特殊原因要 pin 旧版本。老用户升级:

npx skills update hertzflow

通过 vercel-labs/skills universal installer 分发。

1. Surf 账户。在 https://agents.asksurf.ai/ 注册免费账户, 按 官方文档 装 Surf CLI, 然后:

   surf auth --api-key sk-...        # 持久化
   # 或
   export SURF_API_KEY=sk-...        # 环境变量

2. curl 和 jq。macOS / Linux 自带。Windows 用 WSL 或 Git Bash。

匿名 Surf tier 跑不完整份报告 — 注册免费账户拿赠送 credit 后才够用。

CA + Alpha 上线日期
        ↓
forensic_pipeline.py     (Python, ~1-8 分钟, 确定性 — 无 LLM 参与)
  生成 skeleton.json      locked 数据 + <LLM_NARRATIVE_PLACEHOLDER> 槽位
        ↓
LLM 填充                 (你的 agent 只填 ~100-250 个 narrative 槽位)
  生成 filled.json
        ↓
validate_report_data.py  (拦截越界写入 / 幻觉 / 重复填充 / 没 provenance)
        ↓
render_report.py         (Jinja2, 确定性)
  生成 report.md + monitoring/*.json

三层严格的信任边界:

• Pipeline (Python)。占报告 ~70%, 所有 SQL、链上读取、持仓 分布、evidence graph ID 都在这一层。只有 pipeline 接触数据。
• LLM (任意模型)。填 ~100-250 个 narrative 槽位 (取决于代币活跃度
  • m6 谱系大小)。不能写 SQL, 不能选 verdict, 不能定阈值。
• Validator。拦截 LLM 越界写入、引用不在 evidence graph 的 数据、以及重复填充的样板话。

结果: 不同 LLM 跑同一份 skeleton 得到同一个 verdict。跨模型一致 性测试显示 ~91% 叶子节点完全相同, verdict 和 action enum 100% 匹配。

报告支持中英文双语 — 默认中文。forensic_pipeline.py 和 render_report.py 都接受 --lang {zh,en} 参数, 两步必须一致。

python3 skills/hertzflow/alpha/v06/forensic_pipeline.py \
    --ca 0x... --out-dir ./out/ --lang zh

python3 skills/hertzflow/alpha/v06/render_report.py \
    --skeleton ./out/.work/skeleton.json \
    --filled   ./out/.work/filled.json \
    --out      ./out/report.md \
    --lang     zh

out/
├── report.md                              # 用户报告 (Markdown)
├── monitoring/
│   ├── monitoring_paste.json              # Binance / OKX 粘贴导入
│   ├── monitoring_gmgn.txt                # GMGN 粘贴 (纯地址列表)
│   ├── monitoring_gmgn_quoted.txt         # GMGN 粘贴 (带引号格式)
│   ├── monitoring_binance_web3.csv        # Binance Web3 CSV 上传
│   ├── monitoring_okx.csv                 # OKX CSV 上传
│   └── monitoring_wallets_full.json       # 分析用 (完整记录, 含余额 + 派发 %)
└── .work/                                 # 隐藏中间文件
    ├── skeleton.json
    └── filled.json

报告是一份 Markdown, 自顶向下按结构往下看。如果只看一眼, 看最顶上 的 🎯 一屏结论 就够了 — 那是把整份报告浓缩成 7 行的版本。下面 的表格跟你在 render 出的报告里实际看到的章节标题一一对应 (顺序也 按 render 出的顺序排)。

如果报告头部有 ⚠️ N 条文字质量提示 这一行, 说明数据是干净的但 narrative 槽位质量不达标 (重复模板 / 数字幻觉 / 通用 boilerplate), 让 LLM agent 重填那几个槽位即可。 数据本身是 pipeline 锁住的, 不论 warning 在不在都已经过 validator。

CEX 上线 tier 是报告里最重要的上游催化分类 — 它决定你对这个币 该期待什么样的价格走势, 以及出场流动性有多少。

Tier 是从真实上线数据算出来的, 不是从上线时间推测。看报告头部 的 tier 字段。

monitoring/ 文件夹是可执行部分 — 报告告诉你要盯的钱包, 已经按四种主流 Alpha 交易者用的 tracker 格式导出好了。粘贴 导入比文件上传可靠, 优先用粘贴。

1. 打开 Binance Wallet App → 地址追踪 / Track address。
2. 选粘贴导入 JSON (Binance 界面叫 "GMGN 格式")。
3. 打开 monitoring/monitoring_paste.json, 全选复制, 粘进对话框。 每行是 {address, name, emoji}, name 含角色标签 (例如 H-Deployer-3a6dc)。
4. 确认 — 地址带角色标签出现在监控列表里。

粘贴失败就走文件上传, 用 monitoring_binance_web3.csv (Binance Web3 钱包文件上传只接受 CSV, 不接受 JSON)。

1. OKX Wallet → Watchlist / Track address。
2. 选粘贴导入。同一份 monitoring_paste.json 直接用。
3. 粘贴失败改 monitoring_okx.csv (列: Network, Address, Label, Note)。

GMGN 的 bulk import widget 当前疑似只支持 Solana — 我们测试 BSC 0x… 地址粘贴都报 "Invalid format"。如果你那边 GMGN BSC bulk import 能用, 先试 monitoring_gmgn.txt (纯地址列表), 不行再试 monitoring_gmgn_quoted.txt (带引号数组)。都不行就 GMGN UI 里 一个一个手动加, 或者干脆只用 Binance / OKX。

这是完整记录集 — pipeline 识别到的所有钱包, 含当前余额、派发 %、 角色 enum、evidence graph 引用、近 72h 活动 flag。不能导入任何 tracker UI, 用来自己写脚本告警或灌进表格分析。

钱包标签格式 <SYM>-<ROLE>-<addr5> (例如 H-Deployer-3a6dc / H-Dumper-1de14), ≤25 字符跨平台兼容。角色 enum: Deployer / Dumper / PDumper / Quiet / LP / Anomaly / Other。

MIT — 自由使用、修改、再分发。注明出处友好但非必须。

由 HertzFlow 构建并开源回馈社区。 HertzFlow 是 BNB Chain 上的 permissionless perpetuals 交易所, 这个 skill 是我们内部使用的开源工具套件之一。