wbdou2006/ClaudeSkills

Stock analysis

This skill produces the same report every time:

• an overview table
• one panel per ticker, containing:
  • facts
  • a weekly chart against the 52-week high
  • returns
  • the drivers of the drop
  • the bull and bear cases
  • holdings, data gaps and the calculation log

The scripts do all the arithmetic and the HTML. Your job is the research and the checking. Do not redesign the page or hand-write HTML.

SKILL_DIR is the folder containing this file. Find it once with dirname on this SKILL.md path.

Files

Path	Purpose
scripts/compute_drawdown.py	daily CSV → 52-week high and low, % drop, returns, weekly closes
scripts/verify.py	re-derives every headline number from the CSVs, checks fields and sources, writes verified.json, exits 1 on any problem
scripts/build_report.py	worker JSON + verified.json → one self-contained HTML page (uses assets/report.css)
scripts/build_email.py	same verified numbers → ready-to-send drawdown_report.eml (email-safe HTML + text, report attached); validates the address
scripts/send_email.py	for the user to run on their own computer: sends the .eml over SMTP with a confirmation prompt and a locally entered app password
defaults.json	the default watchlist used when the user names no tickers (user-editable)
references/worker_brief.md	the brief each worker follows (fill the {…} placeholders)
references/schema.md	the <TICKER>.json contract
examples/run-2026-10-16/	a complete reference run (FTEC, QQQ, SMH, ARTY): four worker JSONs and the finished report.html. Use it as the bar for depth, tone and citation style.

Workflow

1. Set up

1. Read the ticker list from the user's request.
   • No tickers named ("run my stock analysis", "how's the market vs highs?"): read $SKILL_DIR/defaults.json and use its default_watchlist, in that order. Say in your first line that you're using the default watchlist and list it.
   • Tickers named: use exactly those; don't add the defaults unless the user asks (e.g. "the usual plus NVDA").
   • Indexes (S&P 500, Dow Jones, Nasdaq Composite…): research the index level itself, using the id from defaults.json (SPX, DJI) or a short label like COMP as the ticker in file names. Tell the worker it's an index (see the brief's index notes).
2. Create a work folder: WORK_DIR=$PWD/drawdown-<YYYY-MM-DD>.
3. Create a task list with:
   • one task per ticker
   • "Verify numbers"
   • "Build and publish report"

2. Fan out

Spawn one subagent per ticker, in parallel and in the background. Each one's prompt is:

• the full text of references/worker_brief.md,
• with {TICKER}, {WORK_DIR}, {SKILL_DIR} and {TODAY} filled in,
• plus any ticker-specific angles worth checking. For a chip fund, for example: export controls, hyperscaler capex, memory cycle.

When the user names only one or two tickers and has not asked for subagents, you may do the research yourself by following the same brief.

3. Wait

When a worker finishes, tick its task off and post one line: the ticker, its % drop, and the date of its high. Do not read worker transcripts.

4. Verify

Run:

python3 $SKILL_DIR/scripts/verify.py --dir $WORK_DIR --tickers <T...>

If it reports MISMATCH, MISSING or NO SOURCE/TS, fix the problem before building:

• Arithmetic or hand-copy errors: re-run compute_drawdown.py and correct the JSON from <T>_metrics.json.
• Missing research: send the worker back with SendMessage.

UNCITED and NO CROSS-CHECK are warnings. Mention them in your reply if they remain.

5. Build

Run:

python3 $SKILL_DIR/scripts/build_report.py --dir $WORK_DIR --tickers <T...> \
  --out $WORK_DIR/drawdown_report.html --title "<2-4 word name>" --eyebrow "<e.g. ETF drawdown review>"

Choosing the title: pick a short, specific name, such as "Tech ETF Drawdowns" or "Chip Stock Pullbacks". Use the same title for repeat runs on the same basket.

6. Publish

Publish the HTML with the Artifact tool, using icon: "chart" and a one-sentence description that names the tickers and the as-of date.

• Repeat run on the same basket: if the user wants the old link kept, republish to that artifact's URL.
• Do not edit the generated HTML. Report-wide changes belong in build_report.py or assets/report.css.

7. Email (only when the user gave an address)

Skip this step unless the user's own message includes an email address, e.g. "…and email it to [email protected]". If they say "email it to me" without an address, ask for one; never guess or reuse an address from elsewhere.

No connectors. Do not use Gmail, Outlook or any other email connector or browser automation, and never ask for or handle an email password. Sending happens on the user's side: you prepare a ready-to-send message, the user sends it.

1. Build the email.

   python3 $SKILL_DIR/scripts/build_email.py --dir $WORK_DIR --tickers <T...> --to <address...> \
     --report $WORK_DIR/drawdown_report.html --artifact-url <published link> --title "<same title>"
   

   Exit code 2 means the address is invalid: ask the user to correct it. The script writes drawdown_report.eml (HTML + plain text, report attached, recipients and subject filled in), plus email_body.html, email_body.txt and email.json.
2. Deliver it. Send drawdown_report.eml with SendUserFile (display attach).
3. Tell the user how to send it, in two or three lines:
   • Open the .eml in Outlook (opens as a draft) or Apple Mail (Message → Send Again), check it, and press Send.
   • Or, from Terminal on their own computer: python3 send_email.py drawdown_report.eml --from <their address>. The script shows the recipients, asks for confirmation, and prompts for an app password itself. Send scripts/send_email.py too if they ask for it.
4. Privacy of the link. The artifact is private to the user. If the recipient is someone else, say they'll rely on the attached report unless the user shares the page.

8. Reply

Keep the reply short:

1. One sentence saying what the page is.
2. A markdown table sorted by drop, with columns: ticker, last close, 52-week high (date), drop.
3. Two or three bullets with the main findings across tickers.
4. A short "Data gaps" note covering the tool failures that affected the run.
5. If an email was requested: one line saying the .eml is ready to send and how.

Do not recap the steps.

Non-negotiables

1. Every number comes from a tool call, with its source and its timestamp.
2. "No data" beats a guess.
3. All math is done in the scripts.
4. Read-only tools only. Never touch order, alert or watchlist tools. Never send email yourself; step 7 only prepares a message for the user to send.
5. No buy or sell recommendation. Present both cases fairly.