多资产类别投资多视角分析工作流 (Multi-Asset, Multi-Expert Investment Analysis)

Adding a new class (ETF, commodity, real estate) = drop a new asset_profiles/<class>.md file in the same format; no changes needed to this SKILL.md.

Prerequisites on the host system:

• uv — astral-sh/uv (used to run the Python scripts with PEP 723 inline dependencies). Scripts auto-install their deps on first run.

跨步骤规则：货币单位统一 (Currency-Unit Consistency)

In every file produced by this skill, every monetary value MUST be explicitly labeled with its currency (USD / HKD / CNY / etc.). The report's primary currency (CURRENCY) is determined in Step 1 from the asset profile: stocks → trading currency of the exchange; crypto → USD; private companies → usually USD for US companies, confirm for non-US.

Rules:

1. All prices, market caps, revenue, FCF, target prices, valuations, funding rounds, TVL, etc. must carry an explicit currency suffix — e.g. 175.23 USD, ¥2,145 亿 CNY, HK$ 420.5, 95,000 USD (for BTC).
2. If a source cites a different currency (e.g. a Wall Street estimate for a HK company quoted in USD, or a Chinese private company round quoted in CNY while the report uses USD), convert to the primary currency and note the conversion: 折合 USD ~= 22.40 (汇率 7.82, 2026-04-16).
3. Never mix currencies implicitly in the same table or sentence without explicit labels.
4. Every analyst agent and the supervisor must follow this rule. Pass the primary currency and the requirement explicitly in their prompts.

步骤一：识别投资标的 (Asset Identification)

1a. Detect ASSET_CLASS

Pick one of stock / crypto / private based on the user's question:

• stock — user mentions a ticker pattern (AAPL, NVDA, 00700, 00700.HK, 600519, 600519.SS), names a well-known public company (Apple, Tesla, 腾讯, 茅台), or uses the word "stock" / "股票" / "A股" / "港股" / "美股".
• crypto — user mentions a crypto symbol (BTC, ETH, SOL, DOGE), uses words like "crypto" / "coin" / "token" / "加密货币" / "币" / "链" / "DeFi" / "L2", or references an L1/L2 protocol.
• private — user mentions a known privately held company (Stripe, OpenAI, SpaceX, xAI, Anthropic, 字节跳动/ByteDance, SHEIN), uses phrases like "private company" / "pre-IPO" / "startup" / "unicorn" / "未上市" / "一级市场" / "估值", or names a company with no public ticker.

If ambiguous (e.g. the user just says "analyze ETH" — could be Ethereum the coin or a less-famous namesake; or says "analyze Palantir" — Palantir was private until 2020 but the user may mean the public stock today), ask the user to clarify before proceeding.

1b. Load the asset profile

Read the matching profile file:

<skill-dir>/asset_profiles/<ASSET_CLASS>.md

Where <skill-dir> is typically ~/.claude/skills/analyze-asset-with-agents-team/. Treat the profile's ## identification section as authoritative for how to resolve the asset.

1c. Resolve and record asset variables

Following the profile's ## identification guidance, use WebSearch to confirm the asset and fill in this unified variable set that every downstream step reads:

• ASSET_CLASS — stock / crypto / private
• ASSET_ID — canonical identifier (AAPL, 00700.HK, 600519.SS, BTC, ETH, stripe, openai)
• ASSET_NAME — full human-readable name
• ASSET_VENUE — exchange / chain / domicile (see profile)
• CURRENCY — primary reporting currency
• Any class-specific helper vars defined in the profile:
  • stock → TICKER, MARKET (US / HK / A)
  • crypto → COIN_ID (CoinGecko), SYMBOL
  • private → no helper vars; valuation data comes from WebSearch only

If multiple candidates match the user's input, ask for clarification before proceeding.

Note: CURRENT_PRICE is not set here — it is fetched by the profile's data script in Step 3 (for stock and crypto) or by WebSearch from primary funding sources (for private).

步骤二：时间戳、工具/模型标识、BASE_DIR

2a. 获取时间戳 (UTC+8)

TZ='Asia/Shanghai' date '+%Y-%m-%d_%H:%M:%S'

Record as TIMESTAMP (for filenames, e.g. 2026-04-17_10:45:00). Also:

TZ='Asia/Shanghai' date '+%Y-%m-%d %H:%M:%S'

Record as DISPLAY_TIMESTAMP (for file contents).

2b. 确定 TOOL_NAME 和 MODEL_NAME

These tag the output directory so the user can tell which CLI + model produced the analysis.

• TOOL_NAME: the CLI tool running this skill. Use a filesystem-safe lowercase slug.
  • Claude Code CLI → claude-code
  • Kimi CLI / Kimi Code → kimi-code
  • OpenAI Codex CLI → codex-cli
  • Unknown → default
• MODEL_NAME: the model identifier. Use a filesystem-safe lowercase slug.
  • Examples: sonnet-4.6, opus-4.6, haiku-4.5, k-2.5, gpt-5-codex
  • Unknown → default

Sanitize both: lowercase, keep only [a-z0-9.-], replace other chars with -, collapse repeats.

2c. 生成 QUERY_SLUG

Short filesystem-safe English/pinyin slug derived from the user's question: core keywords only, [a-z0-9-], ≤40 chars.

2d. BASE_DIR

BASE_DIR = {TIMESTAMP}_{QUERY_SLUG}_{TOOL_NAME}_{MODEL_NAME}

Examples:

• 2026-04-17_10:45:00_analyze-apple-aapl_claude-code_sonnet-4.6
• 2026-04-17_10:45:00_analyze-btc_claude-code_sonnet-4.6
• 2026-04-17_10:45:00_analyze-stripe_claude-code_sonnet-4.6

Create the directory tree:

BASE_DIR="{TIMESTAMP}_{QUERY_SLUG}_{TOOL_NAME}_{MODEL_NAME}"
mkdir -p "${BASE_DIR}/01_basic_data/assets" \
         "${BASE_DIR}/02_extra_data" \
         "${BASE_DIR}/03_answers" \
         "${BASE_DIR}/04_summary"

All subsequent paths are relative to {BASE_DIR}/. For ASSET_CLASS = private, the 01_basic_data/assets/ directory may end up empty (no charts are generated); leave it in place for consistency.

步骤三：获取价格与走势图（Profile-Driven Data Fetch）

Look up the ## data_fetch_command section in the asset profile loaded in Step 1b and execute it. The profile tells you which script to run (if any), what arguments to pass, and which JSON fields to capture.

• stock → runs scripts/fetch_stock_data.py (akshare → yfinance fallback) and writes 3Y daily + weekly K-line PNGs.
• crypto → runs scripts/fetch_crypto_data.py (CoinGecko → Binance fallback) and writes 3Y daily + weekly candlestick PNGs.
• private → no script is run. Fill CURRENT_PRICE / PRICE_AS_OF / data_source from WebSearch on primary funding rounds per the profile's ## data_fetch_command instructions.

Standard JSON fields returned by the fetch scripts (both stock and crypto use the same schema):

• current_price → CURRENT_PRICE
• as_of_date → PRICE_AS_OF
• daily_chart_relpath, weekly_chart_relpath → paths used inside the markdown
• daily_trend_description, weekly_trend_description → pre-generated trend text
• data_source → akshare / yfinance / coingecko / binance
• period_start, period_end, period_high, period_low — all in CURRENCY

If the script exits non-zero or returns {"error": ...}, fall back to WebSearch for the current price and skip chart embedding, but continue the rest of the workflow. Note the degradation in the report.

步骤四：收集 MECE 基础数据 (Profile-Driven WebSearch)

Open the asset profile loaded in Step 1b and run an independent WebSearch (or WebFetch for detail pages) per item in these three sections:

• ## macro_queries — external environment (rates, liquidity, sector health, regulation)
• ## fundamental_queries — what the asset is worth intrinsically (earnings/moat for stocks; tokenomics + on-chain for crypto; funding history + business metrics for private)
• ## market_queries — how the market is currently priced and positioned (valuation multiples, flows, sentiment)

Substitute placeholder variables (<ASSET_ID>, <ASSET_NAME>, <industry>, <SYMBOL>, etc.) from the Step 1c variable set before issuing each query.

Tag every monetary value you extract with its original currency, and convert to CURRENCY where needed (see the currency-consistency rule at the top).

步骤五：保存基础数据文件 (Profile-Driven Template)

Find the ## basic_data_template section in the asset profile. That markdown block is your output template. Fill in the Step 1c variables + Step 3 fetch results + Step 4 WebSearch findings, then save to:

{BASE_DIR}/01_basic_data/basic_data.md

(BASE_DIR already encodes {TIMESTAMP}_{QUERY_SLUG}_{TOOL_NAME}_{MODEL_NAME}, so inner filenames stay short.)

For ASSET_CLASS = stock and ASSET_CLASS = crypto, this includes the two K-line / candlestick PNGs from Step 3. For ASSET_CLASS = private, the template replaces charts with a funding-timeline table — no image embedding is expected.

步骤六：并行派发 6 个分析师 Agent

⚠️ CRITICAL: Dispatch all 6 analyst agents IN PARALLEL in a single message (multiple Task tool calls in one response). Do NOT wait for one to finish before the next.

How to dispatch each analyst

For each of the 6 analysts:

1. Read <skill-dir>/agents/<analyst-name>.md. The skill dir is typically ~/.claude/skills/analyze-asset-with-agents-team/.
2. Strip the frontmatter (between the first --- and the second ---). Keep the body as the persona prompt.
3. Launch a general-purpose Task agent with a prompt containing:
   • The full persona body as role/system context
   • The user's original question
   • Basic data file path: {BASE_DIR}/01_basic_data/basic_data.md
   • Asset metadata (class-aware):
     • ASSET_CLASS — stock / crypto / private (always include — agents will adapt their framework based on this)
     • ASSET_ID, ASSET_NAME, ASSET_VENUE
     • CURRENT_PRICE {CURRENCY} (for private companies pass "last-round implied price {CURRENCY}" or "N/A" with the latest post-money valuation and date)
     • PRICE_AS_OF, TIMESTAMP
   • Currency rule (restate explicitly): primary currency is {CURRENCY}; every monetary value must be explicitly labeled with currency; non-primary currencies must be converted with a noted exchange rate.
   • Output paths (use the analyst short name, i.e. filename stem of the agent .md):
     • Extra research data → {BASE_DIR}/02_extra_data/<analyst-name>.md (e.g. buffett-analyst.md)
     • Investment opinion → {BASE_DIR}/03_answers/<analyst-name>.md

The 6 analysts (dispatch all in parallel)

Agents will adapt to ASSET_CLASS themselves (see the "Asset-class adaptation" paragraph at the top of each agent body). For crypto and private companies, expect some analysts to explicitly note "outside my circle of competence" or "my framework barely applies here" — that is valuable signal for the supervisor, not a failure mode.

步骤七：派发 Supervisor Agent

Once all 6 analyst agents complete:

1. Read <skill-dir>/agents/supervisor-analyst.md, strip frontmatter.
2. Launch one general-purpose Task agent with:
   • The supervisor persona body
   • User's original question
   • ASSET_CLASS, ASSET_ID, ASSET_NAME, ASSET_VENUE, CURRENT_PRICE {CURRENCY} (or "N/A — last-round valuation" for private), PRICE_AS_OF
   • Currency rule (restated): primary currency {CURRENCY}; all money values explicitly labeled and consistent.
   • Paths to all 6 analyst answer files in {BASE_DIR}/03_answers/ (e.g. buffett-analyst.md, munger-analyst.md, …)
   • Timestamp TIMESTAMP, tool TOOL_NAME, model MODEL_NAME
   • Output: {BASE_DIR}/04_summary/summary.md
   • Reminder to handle missing perspectives gracefully: for cryptos and private companies, some technical / market-flow sections may be absent or "N/A"; that is expected and should be noted rather than filled with speculation.

步骤八：展示最终总结

Read {BASE_DIR}/04_summary/summary.md and display the full contents to the user.

错误处理

• If an analyst agent fails, record the error, continue, and note the missing perspective in the final summary.
• If the Python data script fails (e.g. network issue), fall back to WebSearch-based current price and skip chart embedding; note the degraded state in the report.
• If WebSearch returns no data for a sub-item, fill with 「数据暂时无法获取，建议查阅专业金融数据库」.
• If the user's input is ambiguous (e.g. asset class unclear, or multiple assets share the same name), confirm identity first before proceeding.
• If the user's asset doesn't cleanly fit any supported class (ETF, commodity futures, real-estate fund, etc.), tell them which classes are currently supported and ask how they'd like to proceed — do not silently bucket it into a wrong class.

Disclaimer

This skill produces AI-generated analysis for informational purposes only. It is not investment advice. Users should conduct their own due diligence and consult a qualified financial advisor before any investment decision.