AlphaDesk — Investment Portfolio Daily Analyzer

Hand it a screenshot, CSV, or broker MCP feed. It hands back a 9-section daily report with concrete P0/P1/P2 actions, risk dashboard, and a devil's-advocate review — in under 6 minutes for the standard mode.

Who this is for

You have a brokerage account (FUTU / Moomoo / IBKR / Robinhood / 雪球 / 老虎 / 富途 / etc.). You want a professional second opinion every day without hiring an analyst — concrete reasoning about what to hold, trim, hedge, or ignore, with explicit risk numbers and an explicit devil's-advocate take.

This skill gives Claude the structure to do that consistently — same 9 sections every report, same red-line guardrails, same anti-bloat protections so the framework itself doesn't become a governance pet project.

What you get

▍0  算法准确率回顾                  (Did yesterday's call work?)
▍1  Executive Summary               (3-sentence top-of-mind)
▍2  Regime + 宏观仪表盘              (HSI / SPY / sector temperature)
▍3  持仓快照                        (positions table)
▍4  重仓股深度研究                   (1-3 stocks deep, depending on mode)
▍5  推荐标的研究                     (only if action triggered)
▍6  Smart Money + 资金流             (institutional vs retail flow)
▍7  风控仪表盘                       (concentration / leverage / risk level)
▍8  P0 / P1 / P2 执行策略            (with exact qty + reason + 24h enforcement)
▍9  事件日历 + 反向推演               (calendar + Devil's Advocate)

Three modes — picked by your phrasing

Mode	Triggers	Tool calls	Time	When to use
Quick	"快速日报" / "晨报" / "给我看看"	≤10	1-2 min	Daily morning glance
Standard (default)	"今日日报" / "详细" / unspecified	18-30	4-6 min	End-of-day comprehensive
Deep	"深度日报" / "重仓深度" / "财报深度" / "月度复盘"	30-50	10-15 min	Earnings / major news / monthly review

Three ways to feed your data — pick one

Option A — Screenshot (zero setup, fastest)

Send Claude your broker app's holdings screenshot. Claude reads it, structures the positions, and runs the 9-section analysis. Best for first-time use.

Option B — CSV export (most reproducible)

Most brokers and trackers (雪球, 富途, IBKR Flex Query, Robinhood export) produce a CSV with at minimum: code, name, qty, avg_cost, market_value. Drop it at $ALPHADESK_ROOT/personal-holdings.csv and Claude will pick it up each session.

Option C — Broker MCP (live, automated)

If your broker exposes an MCP (FUTU OpenD, Moomoo, others), connect it once and Claude pulls live positions / cash / leverage every session. See docs/data-sources.md.

You do NOT need to choose all three or fill out 7 config files. One data source is enough. The advanced templates in templates/advanced/ are for users who actually have multiple brokers + cross-account reconciliation needs — most users skip them.

Quick start (single-account, 5 minutes)

# 1. Pick a project root
export ALPHADESK_ROOT="$HOME/Desktop/my-portfolio"
mkdir -p "$ALPHADESK_ROOT" && cd "$ALPHADESK_ROOT"

# 2. Install minimal scaffolding
cp -r ~/Desktop/alphadesk-skill/templates/basic/* "$ALPHADESK_ROOT/"
cp -r ~/Desktop/alphadesk-skill/tools "$ALPHADESK_ROOT/"

# 3. Initialize git baseline (so you can roll back any decision)
git init -b main && git add -A && git commit -m "baseline" && git tag baseline

# 4. Drop in your data — pick ONE:
#    A) screenshot.png in the directory
#    B) personal-holdings.csv in the directory
#    C) broker MCP (separate setup)

# 5. Open Claude, say:
#    "今日日报"  (Standard mode)
#    or paste your screenshot

Claude will run the session-start protocol automatically and generate daily-reports/YYYY-MM/daily_YYYYMMDD.html.

Hard constraints (red lines)

These are not made-up. Each comes from a real failure in the private predecessor. See docs/lessons-learned.md for the case files.

Red line	Why it exists
Every portfolio-level metric declares its data source	Concentration / leverage / VaR computed from broker view ≠ from personal CSV. Mixing them produces fictional numbers.
Report gate runs BEFORE HTML write, not after	A gate that runs after the file exists is theater. Schema fail = abort, not "fix later".
P0 decisions have 24h forced-execution rule	Either execute (real or sim) or explicitly cancel. Otherwise next day's report banner turns red. Prevents "P0 monitoring" sham.
No new framework layer without ≥3 evidence samples	The single biggest failure mode: each "report bug" triggers a new red-line / new tool / new protocol. After 8 months it's 64% governance. Don't.
Session breadcrumbs on disk, not in chat	Conversation summaries lie. `tools/sstate.py` writes a journal that survives context compaction, new chats, even Claude restarts.
No emotional language in reports	Words like "便宜" / "贵" / "still losing money" don't belong. State the number, state the rule, state the action.

Session-start protocol (Claude runs this automatically)

# Disk-truth check (before any analysis)
python3 tools/sstate.py show          # what was just done?
python3 tools/sstate.py inventory     # what reports exist today?

# Phase progress (only if you're tracking evidence-first)
python3 tools/evidence_log.py status

This recovers state even if your previous chat ended hours ago.

Sub-documents

Doc	What it covers
`README.md`	Full user guide: install → daily use → customization
`docs/data-sources.md`	Screenshot / CSV / MCP — how to plug in your data
`docs/workflow.md`	The 9 sections in detail + report_gate flow
`docs/architecture.md`	Pipeline (data → reconciliation → gate → render → log)
`docs/governance.md`	Why this skill stays small + anti-bloat self-check
`docs/lessons-learned.md`	Failure cases that drove each red line
`docs/advanced.md`	Multi-account reconciliation, friend-held collateral isolation, three-book contract

When NOT to use this

You want a stock-picking bot — this is for analysis, not auto-trading. Final decision is yours; this skill never places real orders without your explicit confirmation.
You want backtesting / quant research — use vectorbt / backtrader / qlib. This is daily tracking + decision support.
You only check holdings monthly — overkill. Use a spreadsheet.
You want institutional compliance theater (4-eyes / IPS / Brinson / TCA) — deliberately removed. Single-user mode = those produce paperwork, not alpha. See docs/lessons-learned.md §4.

Triggers — when Claude should auto-load this

Auto-load on:

日报 / 持仓日报 / portfolio report / daily report
持仓分析 / 持仓快照 / holdings analysis
重仓深度 / 月度复盘 / earnings deep-dive
卖一半评估 / 加仓评估 / P0 决策 / position sizing
风控仪表盘 / 集中度 / leverage / 止损线
Smart Money / 资金流向 / institutional flow
Devil's Advocate / 反向推演 / 反方论证

Skip if:

Pure single-stock fundamental research → use a research skill
Trade execution / automation → not this skill's job

Version

AlphaDesk v1.0 (2026-05-28). Distilled from a private predecessor that ran 8 months / 11 versions / 64% governance overhead. The cuts and the reasons are in docs/lessons-learned.md.

alphadesk