desmondc9-agent-skills

A collection of Claude Code skills for document processing, investment analysis, and productivity workflows.

Skills

markdown-to-docx

Converts any Markdown file to a professionally formatted Word document (.docx).

Features:

• Native Word Table of Contents (auto-updates in Word)
• Mermaid diagrams rendered to high-resolution PNG and embedded
• Tables with styled header rows, alternating shading, and borders

Install:

npx skills add desmondc9/agent-skills@markdown-to-docx -g

Prerequisites: pandoc, mmdc (Mermaid CLI), python-docx, chromium-browser

See the skill README for full usage and troubleshooting.

---

analyze-asset-with-agents-team — ⚠️ DEPRECATED

已废弃，不再维护。 保留仅供历史参考，请勿用于新的分析任务。

Runs a comprehensive multi-expert investment analysis across any asset class — publicly listed stocks (US / HK / A-share), cryptocurrencies (BTC, ETH, SOL, any token), or privately held companies (Stripe, OpenAI, SpaceX, 字节跳动) — using 6 parallel analyst personas plus a synthesis supervisor.

Features:

• Auto-detects asset class (stock / crypto / private) from the user's query and loads the matching asset profile (asset_profiles/<class>.md)
• Per-class data fetch:
  • Stocks → 3Y daily + weekly K-line charts via akshare / yfinance + mplfinance
  • Crypto → 3Y daily + weekly candlesticks via CoinGecko + Binance fallback
  • Private companies → funding-timeline table + comparable-company multiples (no chart)
• Gathers MECE baseline data per class: macro context, fundamentals/tokenomics/funding history, and market/on-chain/liquidity signals
• Dispatches 6 analyst agents in parallel, each applying their framework honestly to the given asset class:
  • 🔴 Warren Buffett — value investing, economic moats, margin of safety
  • 🟣 Charlie Munger — mental models, inversion, psychology of misjudgment
  • 🔵 Cathie Wood — disruptive innovation, Wright's Law, 5-year targets
  • 🟢 王煜全 — global tech transfer, China industrial clusters, geopolitics
  • 🟡 招财大牛猫 — A/HK market tactics, technical + fundamental hybrid
  • 🟠 段永平 — stop-doing list, business essence, management integrity
• Supervisor agent synthesizes all six perspectives into a structured final report, with graceful handling of class-specific gaps (e.g. technicals are "N/A" for private companies)
• Every run saves to a timestamped folder: 01_basic_data/, 02_extra_data/, 03_answers/, 04_summary/
• Extensible: adding a 4th asset class (ETF, commodity, real estate) is a single new file in asset_profiles/

Install:

npx skills add desmondc9/agent-skills@analyze-asset-with-agents-team -g

Usage: Just ask naturally — the skill auto-triggers on investment-analysis intent across any supported asset class:

# Stocks
帮我分析一下苹果公司值不值得买
analyze Tesla stock for me
腾讯现在能买吗？
Should I buy NVDA now?

# Crypto
analyze BTC
is ETH a good long-term hold?
SOL 还值得买吗？

# Private companies
Is Stripe a good investment?
OpenAI 的估值合理吗？
analyze SpaceX as a private investment
字节跳动值不值得买

Prerequisites: uv for running the bundled Python data-fetch scripts (stock and crypto fetchers). Otherwise a Claude Code environment with WebSearch, WebFetch, Bash, and Task tools enabled.

See the skill README for the full profile-driven workflow and agent personas.

Disclaimer: AI-generated analysis for informational purposes only. Not investment advice.

---

deep-company-analysis

Generates a structured deep-dive research report on any company, stock, or crypto project from both growth and value investing perspectives. Uses an 8-dimension framework covering industry dynamics, management quality, financials, growth drivers, valuation, policy & geopolitics, technology disruption risks, and a crypto-specific appendix.

Features:

• Executive Summary that identifies the 5 highest-impact dimensions with quantitative evidence and tracking KPIs
• Dual-lens analysis: every dimension examined from both growth-investing and value-investing angles
• Strict output discipline: no "buy/sell/hold" recommendations; all data sourced; speculative claims marked; inapplicable dimensions explicitly labeled
• Competitive benchmarking against peers and industry medians for all key metrics
• 3-year / 5-year / 10-year trend analysis (or actual history if < 10 years)
• Dedicated crypto appendix (tokenomics, on-chain metrics, governance risks, MEV, regulatory status) when analyzing blockchain/Web3 projects
• Per-dimension 200-word summaries + a 12-month "综合观察清单" of trackable catalysts

Install:

npx skills add desmondc9/agent-skills@deep-company-analysis -g

Usage: Just ask naturally — the skill auto-triggers on deep-research intent:

深度分析一下腾讯
帮我调研一下 NVDA
research Apple as a long-term investment
OpenAI 的基本面怎么样
analyze ETH from a fundamentals perspective

Prerequisites: A Claude Code environment with WebSearch and WebFetch tools enabled.

See the skill README for the full workflow and REFERENCE.md for the complete 8-dimension analysis template.

---

transcribing-meeting-recordings

Converts any meeting recording (.mp4 / .mov / .mkv / .wav / .m4a / .mp3) into a cleaned, timestamped SRT transcript with real participant names — especially Chinese / multilingual content. Combines whisperx (faster-whisper + alignment + pyannote diarization) with a Microsoft Teams / Zoom video-frame trick: the active-speaker banner in the recording is read off the frame, letting you map anonymized SPEAKER_XX clusters onto the real names of the people in the call.

Features:

• Single-pass whisperx pipeline: ASR (large-v3) + word-level alignment + pyannote speaker diarization, tuned to fit 8 GB GPUs (--compute_type int8 --batch_size 4)
• Post-processor (make_srt.py) that strips Whisper's YouTube-training hallucinations (请不吝点赞订阅...), spaces CJK/ASCII boundaries (shipment加A → shipment 加 A), smooths short diarization "flake" cues, and merges consecutive same-speaker cues
• Speaker identification via video frames — extract one frame per SPEAKER_XX at a clean utterance, read the Teams / Zoom active-speaker banner, fill in a markdown mapping table; the SRT then carries real names
• Handles pyannote over-segmentation (one person split into 2+ clusters) by collapsing duplicate names during merge
• Per-meeting speakers.md mapping + parent-folder speakers.md roster pattern for cross-meeting name tracking (pyannote labels are not stable across recordings; aliases incl. Whisper Chinese-name homophones are recorded)
• Captures every failure mode encountered in practice — Python 3.14 / torchaudio incompatibility, 3 separate Hugging Face license walls, CUDA OOM at default batch size, Whisper hallucinations during silence — so the next run avoids them

Install:

npx skills add desmondc9/agent-skills@transcribing-meeting-recordings -g

Usage: Drop a recording into a folder, then:

ffmpeg -i meeting.mp4 -vn -ac 1 -ar 16000 -c:a pcm_s16le audio.wav
whisperx audio.wav --model large-v3 --language zh --diarize \
  --diarize_model pyannote/speaker-diarization-3.1 \
  --compute_type int8 --batch_size 4 --output_format json
cp <skill>/templates/speakers.template.md ./speakers.md
python3 <skill>/scripts/make_srt.py   # → transcript.srt

Then for each SPEAKER_XX, extract a frame (ffmpeg -ss <t> -i meeting.mp4 -frames:v 1 spk_XX.jpg), read the Teams banner, fill in the name in speakers.md, and re-run make_srt.py.

Prerequisites: ffmpeg, Python 3.11 (not 3.14 — torchaudio incompatibility), uv tool install whisperx --python 3.11, an NVIDIA GPU (≥ 6 GB recommended), and acceptance of three pyannote gated repos on Hugging Face (speaker-diarization-3.1, segmentation-3.0, speaker-diarization-community-1).

See the skill README for the full pipeline, prerequisites, and the Common Mistakes table.

---

analyze-value

Estimates the intrinsic value of a company, stock, or cash-flow-generating asset (commercial real estate, 收租物业) using the valuation methodology from 吴军《财商训练课·企业估值》: ignore cost and historical price — value only depends on how much free cash flow the asset will keep producing, and how risky that is. Implements risk-adjusted discounted cash flow (DCF) with a CAPM-style β discount, plus a baseline check for risky fixed-income (理财 / 票据 / 债券) against the risk-free deposit rate.

Features:

• Risk-adjusted DCF: Value = Σ FCFₜ / [1 + r + β·(rm − r)]ᵗ — higher risk (β) discounts harder, lowering valuation
• Free-cash-flow based, not net profit (易造假、要再投入), and cost-blind (固定资产/历史造价/历史股价 don't enter the formula)
• dcf_valuation.py script outputs per-year discounted FCF, ΣDCF, and valuation under two conventions (standard present-value, and 吴军's equivalent-deposit-principal), validated against the article's worked examples
• Expected-return mode for risky fixed income: probability-weighted payoff discounted back and compared to the bank baseline
• Commercial-real-estate guidance (supply/demand already lives in the FCF; self-occupied homes excluded)
• Discipline: gives a valuation range with β / growth sensitivity, labels currency and assumptions, no buy/sell calls — 追求方向性正确，不追求精确

Install:

npx skills add desmondc9/agent-skills@analyze-value -g

Usage: Ask naturally — the skill auto-triggers on valuation intent:

这家公司值多少钱？帮我估个值
用 DCF 给宁德时代估值
NVDA 现在是高估还是低估？
帮我看看这个商铺值不值这个价
这份年化 5% 的理财，相对存银行到底值多少？

Prerequisites: Python 3 (standard library only — no extra packages). WebSearch / WebFetch recommended for gathering the cash-flow inputs.

See the skill README for the workflow and REFERENCE.md for the full method, formulas, and the article's worked examples.

---

analyze-dip

Judges whether a sharp crash / big pullback in a stock, cryptocurrency, or other asset is a buy-the-dip opportunity or a time to exit, using the 10-step framework from the video 《AI硬件股暴跌，现在该抄底还是逃命？我的10步判断框架》 (Micron as the running case). The core discipline: never judge by how much it dropped or whether the news sounds good/bad — distinguish 「跌的是情绪」 from 「跌的是基本面/盈利逻辑」.

Features:

• 10-question checklist, one key question per step: earnings-logic revisions (EPS revision), multi-model fair value (DCF + multiples), collective analyst re-pricing, growth quality (revenue/EPS growth, margins, key businesses like HBM), free-cash-flow trend, industry-level change, moat depth, balance-sheet safety, fundamentals-vs-sentiment, and the final empty-position test ("if I held none today, would I buy at this price?")
• Gate logic: if step 1 (earnings logic) fails, the rest don't matter — the framework says exit, not dip-buy
• Crypto adaptation table mapping each step's stock metrics to on-chain/protocol equivalents (fee revenue, NVT, MVRV, treasury runway, token unlocks)
• Structured output: a 10-step evidence matrix (✅/⚠️/❌) plus a directional verdict — 偏抄底（情绪错杀）/ 偏离场（逻辑已变）/ 证据不足 — with sources and assumptions labeled
• Doubles as a pre-purchase checklist before long-term holding (the original author's own usage)
• Platform-agnostic data gathering (Investing Pro, filings, WebSearch); no target prices, no investment advice

Install:

npx skills add desmondc9/agent-skills@analyze-dip -g

Usage: Ask naturally — the skill auto-triggers on crash/pullback decision intent:

NVDA 跌了 20%，现在该抄底还是离场？
BTC 暴跌，是抄底机会还是该跑？
帮我看看美光这次回调要不要补仓
buy the dip or get out of TSLA?

Prerequisites: A Claude Code environment with WebSearch / WebFetch for gathering earnings estimates, valuation data, analyst revisions, and sector news.

See the skill README for the workflow and REFERENCE.md for the per-step criteria, historical panic-selloff cases, and the crypto adaptation table.

---

fetch-dedao-cources

Scrapes paid Dedao (得到APP, dedao.cn) courses into a complete local HTML archive — one folder per article (HTML + localized images), plus a searchable index.html master index. Battle-tested on 万维钢's 7 courses (2,115 articles) and 卓克's 7 courses (2,019 articles) with zero misses.

Features:

• Reuses the local Chrome login state (copies Cookies into a headless debug instance — no password needed)
• Full article-list pagination via the site's internal API (count cross-checked against the official course counter)
• Per-article content harvested by navigating real pages and capturing the ddarticle responses (replayed in-page fetches are server-side withheld)
• Images downloaded locally and inserted at their original positions; per-article title image (header_*) preserved; filename-collision protection via URL hashing
• Two-tier promo-image filtering: wide-banner rule (aspect ratio ≥ 3, no caption) + byte-identical repeat-group triage (OCR + QR-code detection + visual review), with a shipped blacklist of 12 verified campaign images
• Configurable via config.json: course list (name + URL), archive-root naming rule ({date}/{year}/{month}/{author} templates), and per-article folder naming rule ({course}/{module}/{date}/{number}/{title} templates — e.g. 万维钢-style {course}-{module}-{date}-{title} or 卓克-style {course}-{date}-{number}_{title})
• Resumable pipeline, built-in audit (list counts, harvest completeness, image refs, index links, ad residue)

Install:

npx skills add desmondc9/agent-skills@fetch-dedao-cources -g

Usage: Ask naturally with course URLs (requires being logged into dedao.cn in local Chrome):

帮我抓取得到课程 https://www.dedao.cn/course/detail?id=... 的全部文章
把卓克科技参考1-5存到本地, 目录按 课程名-日期-编号_标题 命名

Prerequisites: google-chrome, uv, and a logged-in dedao.cn session in the local Chrome. Optional: tesseract + chi_sim traineddata for OCR-based ad triage.

See the skill README for the six-step workflow, config spec, and the critical-gotchas checklist.

---

generate-report

Writes, rewrites, or reviews a research report to the standards of the top consulting and research houses. The methodology is distilled from a side-by-side reading of real 2025–2026 reports by McKinsey, Gartner, BCG, Bain, Deloitte, PwC, IDC, Forrester, WEF, and the IMF/World Bank.

Features:

• Stage 0 gate — three questions before writing: the one-sentence thesis (must work as the subtitle), the counterintuitive hook (the "everyone thinks X, the data says Y" gap), and where every number comes from. Can't answer them → don't write yet, go back to research
• 9-block skeleton (title+subtitle → executive summary → why now → methodology → 3–8 finding chapters → forecasts → role-split recommendations → risks → sources/disclaimer), with three chapter-cut patterns (issue-parallel / value-chain / overview-plus-deep-dives) and three size tiers (1–2 page brief, 8–15 page standard, 30+ page flagship)
• Five-beat per-chapter template: What → What's changing → Why → So what / who wins → Now what
• 10 language & evidence rules: conclusion-first at report/chapter/paragraph level, Action Titles (every heading and exhibit title is a full judgment sentence), one argument per paragraph, a number behind every claim, forecasts with time anchor + figure + baseline (Gartner style), the exhibit trio (number + conclusion title + Sources/Notes), named cases, one coinable framework (BCG's "10-20-70" style), stated uncertainty, role-split recommendations
• Institution playbooks (PLAYBOOKS.md) — each house's structure, data base, exhibit conventions, and imitation notes, plus a selection guide for when the user hasn't named one
• Two review modes: a 10-item release checklist, and a 100-point rubric (10 dimensions × 10 points) with deduction rules and grade bands for scoring someone else's draft
• Reusable genre template — multi-country × direction scoring × value chain × company investment map (templates/): for the "research where China and the US will each push over the next 3–10 years, score every direction 0–100 and rank them, break each one down into upstream/midstream/downstream, then analyze every beneficiary company" class of request. Ships the five-stage pipeline (parallel research → market-data collection → scoring → per-chapter writing → docx), the five-dimension weighted scoring model with anchor table and tier bands, the 9-chapter skeleton, the four-part direction section, the six-part company card (weekly K-line + metrics table + earnings table + three-window (2026-28 / 28-32 / 32-35) bull/base/bear projection + verdict + risks), the five standard tables (scores, value chain, path comparison, valuation tiers, scenarios), per-chapter footnote prefixes, 10 genre-specific quality rules (every high score needs a counter-evidence; official vs third-party-tracked figures labeled separately; conflicting ranges listed not averaged; earnings quality split into operating / one-off / deal-pulse), and 8 known traps
• Market-data script (scripts/fetch_market_data.py): one command pulls ~1.5y weekly K-line PNGs (red-up/green-down by default), a 10-metric stock table, and a two-period earnings table for a whole basket of US / HK / A-share tickers — emitted as paste-ready Markdown (company_metrics.md). yfinance primary, akshare fallback, sparse HK quarters handled, missing data flagged rather than guessed
• Data-integrity red lines: no invented statistics, sample sizes, or citations; every second-hand number carries source + retrieval date; unsourced forward numbers are labeled 【情景假设】
• Delivers into reports/YYYY-MM-DD-<topic>/ (brief.md → report.md → sources.md), and hands off to markdown-to-docx for Word delivery

Install:

npx skills add desmondc9/agent-skills@generate-report -g

Usage: Ask naturally — the skill auto-triggers on report-writing intent:

帮我写一份 2026 年中国 AI 基础设施行业研究报告
把这几份访谈和数据整理成一篇趋势报告，像 Gartner 那样写
这篇报告不够专业，帮我改结构
评审一下这份白皮书，打个分并给修改清单

# 命中投资图谱体裁模板：
帮我调研未来 3-10 年 AI 应用会渗透到哪些领域，美国和中国分别往哪些方向发展，
每个方向 0-100 打分并按可能性排序；各方向的产业链上中下游有哪些公司受益，
逐一分析财报、行业地位、产品线、研发、市值，未来 3-10 年怎么变，值不值得投资，
上市公司附近 1.5 年周 K 线图和指标表格

Prerequisites: WebSearch / WebFetch for sourcing data (or user-supplied material). Optional: the markdown-to-docx skill for .docx delivery; uv + network access to Yahoo Finance / akshare for scripts/fetch_market_data.py (behind the GFW, export the proxy env vars first).

See the skill README for the four-stage workflow, REFERENCE.md for the 12 building blocks, 8 common laws, exhibit/methodology specs, the scoring rubric, and the anti-pattern list, and templates/two-country-scoring-investment-map.md for the investment-map genre.

---

License

Apache 2.0 — see LICENSE.