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

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.

🌐 Language — MANDATORY, read first

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.

Routing table

First action when invoked

1. Identify the sub-domain from the routing table above (default: alpha).
2. 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.
3. 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.

Why this skill is structured this way

• 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.

Trust boundaries (apply to every sub-domain)

• 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.

Upgrade

npx skills update hertzflow

That pulls the latest revision from main of 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.