The hunt prompt

This is the product. It is kept in plain English on purpose so it can be read and edited without touching any code. The scheduled Claude Code session reads this file each morning and follows it.

Two things it produces: the daily screen, and a check on each watchlist ticker.

Part 1 — the screen

Search the web for current data. Do not answer from memory.

Look for stocks or ETFs that have fallen 10% or more in a single day or over one week, within the last 7 days.

Then discard the junk. Skip anything trading under $1 a share, anything below $500 million market cap, shell companies, companies listed within the last year, and biotechs whose value rests on a single drug trial.

From what survives, return the 8 most interesting, favouring those that are also cheap on the numbers — low forward P/E and low price-to-book.

Screen broadly here and do not apply any tighter filters. The phone narrows this list down to his personal thresholds locally, so returning more than he will see is deliberate — it means changing a setting takes effect instantly instead of waiting for tomorrow’s run.

For each one, judge whether the fall is an overreaction or a genuine impairment. The distinction that matters is whether the cause is a one-off event, a fixable problem with a credible timeline, or a permanent change to the business.

Part 2 — the watchlist

Read watchlist.json in this folder. For each ticker listed there, look it up and produce the same tearsheet, whatever its size, sector, or recent price action. Do not filter these out and do not substitute a different company — if he explicitly asked about a stock, answer about that stock.

Accuracy

• Market data more than a day old is worse than useless here. Search for it.
• If you cannot confirm a number, say so. Use null for any price you cannot confirm and "n/a" for any metric you cannot confirm. Never invent a figure. A wrong P/E is a much worse failure than a missing one.

Output

Write the result to brief.json as a single JSON object, nothing else:

{
  "generatedAt": "2026-08-02T10:30:00Z",
  "brief": {
    "asOf": "2 August 2026",
    "note": "<=20 words on what is driving drops right now",
    "candidates": [ ... ],
    "watch": [ ... ]
  }
}

candidates is Part 1, watch is Part 2. Both are arrays of the same shape:

{
  "ticker": "LKQ",
  "name": "LKQ Corporation",
  "dropPct": -15.8,
  "dropWindow": "one day",
  "price": 22.73,
  "prevClose": 26.39,
  "low52": 21.90,
  "high52": 45.20,
  "verdict": "real bargain",
  "business": "<=14 words on what it actually does",
  "cause": "<=30 words on exactly what caused the fall",
  "causeType": "one-off",
  "sector": "Consumer discretionary",
  "marketCapUsd": 5900000000,
  "fwdPe": 8.3,
  "paysDividend": true,
  "metrics": [
    {"label": "Fwd P/E", "value": "8.3x"},
    {"label": "P/B", "value": "1.1x"},
    {"label": "Yield", "value": "2.4%"},
    {"label": "Free cash flow", "value": "Negative"}
  ],
  "right": "<=18 words on the one thing that must go right",
  "wrong": "<=18 words on the one thing that could go wrong"
}

Rules:

• verdict must be exactly one of: "real bargain", "wait and see", "value trap".
• causeType must be exactly one of: "one-off", "fixable", "permanent".
• Prices, marketCapUsd and fwdPe are plain numbers, never strings.
• sector, marketCapUsd, fwdPe and paysDividend are what the phone filters on. Get them right or use null — a null is treated as “unknown” and is never silently filtered out.
• sector should be one of Biotech, Crypto & digital assets, China-domiciled, Airlines, Energy, Banks when one of those fits, since those are the sectors he can exclude. Otherwise use a plain description.
• generatedAt is the actual UTC time of the run, in ISO 8601.