# Skill: company-deep-dive Generate a learning pack that teaches a learner **one specific company** from scratch — what it does, who runs it, where the money comes from, what could go wrong, and why the usual valuation ratios need caveats *for this company*. The goal is that after the pack, someone who had never heard of the stock can follow a news story or an earnings call about it. This is a prompt for a research-capable AI (one that can browse or has current data). Attach [`docs/PROTOCOL.md`](../docs/PROTOCOL.md). Optionally also attach a template from [`templates/`](templates/) and ask the model to fill it in instead of starting blank — templates fix the structure and ids, so a refreshed pack keeps your progress. --- ``` You are a learning-pack generator for StockLearn. Output ONLY valid JSON conforming to the attached PROTOCOL.md (schemaVersion 1.0), no prose, no markdown fences. Company: [name and ticker, e.g. "The Walt Disney Company (DIS)"] Language: [e.g. "en"] As of: [today's date, ISO] Research the company first (latest annual report / 10-K, the most recent quarterly results, investor-relations materials, and reputable news from the last 12 months). Then write a pack with EXACTLY these five units, in this order, using these unit ids: 1. "overview" — what the company is and how it makes money. Business model (how the parts reinforce each other), the reporting segments as the company itself defines them, and the key history (founding, transformative acquisitions, pivots). 2. "leadership" — CEO and tenure, recent or planned succession, board/governance points a newcomer should know, notable past capital- allocation decisions. Include how long the current CEO has been in the role and what the market considers their track record. 3. "financials" — revenue by segment, operating income by segment (as a markdown table with the fiscal period named), key balance-sheet items (cash, debt, goodwill, shareholders' equity), and 3-year trend for revenue and operating margin. 4. "headwinds" — current legal/regulatory matters, structural pressures on the business (e.g. declining channels, cost inflation, competition), and execution risks management itself has flagged. 5. "valuation-caveats" — the company's current P/B, P/E, EV/EBITDA (or whatever ratios are most used for this industry) AND, for each, a card explaining why the raw number misleads for THIS company (goodwill from acquisitions, one-off impairments, cyclical earnings, leases, etc.), plus a final "what to watch" card listing metrics and events that would change the picture — stated neutrally, with no buy/sell/hold language. Rules: - 3-5 cards and 4-6 questions per unit. - Separate PRINCIPLE cards (timeless: how the business model works, why goodwill inflates book value) from DATA cards (figures, names, ongoing events). Every DATA card MUST have "dataAsOf" set to the date the facts were true and MUST name the fiscal period or source in the body ("FY2025 10-K", "Q3 FY2026 earnings release", "press release 2026-05-14"). - Tag cards "principle" or "data"; add "leadership", "segments", "risk", "valuation" tags as appropriate. - Question mix: at least 40% must test UNDERSTANDING (why does X matter, what happens to Y if Z) rather than recall. Recall questions (who is the CEO, which segment is largest) are fine but must not dominate. - Numeric questions ask for ranges or rankings ("roughly what share", "which segment is largest"), never exact figures to three decimals. - NEVER ask the learner to plug numbers from the prompt into a formula ("price $217.89, EPS $7.91 — what is the P/E?", "subtract the one-off and recompute"). Anyone who reaches the valuation unit can divide; such questions test arithmetic, not understanding, and insult the learner. Put the computed ratio in the card and ask what it hides, why it moved, or what would change it. An "estimate" question is only justified when the learner must supply the magnitude from memory or judgment (e.g. "roughly what share of monthly users pay?"), never when every input is printed in the question. - Every question sets "cardRefs" to the cards it tests and has a genuine explanation that teaches, not just "the answer is B". - Use stable, descriptive ids (cards: "ceo", "revenue-by-segment"; questions: "q-largest-segment"). If a template is attached, KEEP its ids. - Set pack "id" to "-deep-dive", "topic" to "companies/", "generatedAt" to the As-of date, and "source" to "skills/company-deep-dive.md". - Educational only: never include buy/sell/hold recommendations, price targets, or "undervalued/overvalued" verdicts. Describe, explain, compare. - Write in the requested language. ``` --- ## Refreshing a pack later Company facts rot. To update a pack you have already studied without losing progress: ``` Here is my existing StockLearn pack for . Regenerate it with current data as of : update every DATA card and its dataAsOf, update any recall question whose answer changed, and KEEP every pack/unit/card/question id exactly as it is so my review schedule carries over. Keep PRINCIPLE cards unless they are wrong. Output only the JSON. ``` Re-importing a pack with the same `id` replaces the old one; progress is keyed by ids, so unchanged ids keep their schedules. ## Filling a template Templates in `templates/` are valid packs whose time-sensitive content is marked with `«FILL: …»`. Attach one and add to the prompt: ``` Use the attached template: keep its structure and ids, replace every «FILL: …» marker with researched content, set every dataAsOf, and remove the "(template)" suffix from the title and the template note from the description. ``` Before importing, check that no `«` characters remain — the validator will accept the file either way, so an unfilled marker would simply show up as text on a card.