Deep Dive

The quick kill verdict is a smoke test. A deep dive is the due diligence you do before spending real money or months of work. It follows every rule of ${CLAUDE_PLUGIN_ROOT}/skills/kill/SKILL.md (safety §0, source rules §3, metrics §4, rubric and fixed rules §5) with higher minimums.

Steps

1. Load or create the idea.
   • If ideas/*-<slug>/idea.json and verdict.json exist, read them. Keep the old score, verdict, mode, evaluated and source count: they go into history.
   • Otherwise run sections 1–2 of the kill skill first: the brief extraction and the questions. Country matters even more here.

2. Run 6 research tracks in parallel, one subagent each when available, in the foreground (run_in_background: false when the tool offers it), and wait for all of them.

   • Every track returns its facts with source links (URL, title, publisher, date, tier), its estimates with their formula, and its unknowns.
   • Search in the language of each target country, on local sites.

   Track	Must find
   Market	TAM / SAM / SOM from at least 2 independent sources (1 primary), growth rate, 5-year trend
   Competitor teardown	Top 10 (global + local): offer, price tiers from the pricing page (dated), traction (traffic, reviews, funding, headcount), positioning, weakest point
   Customer voice	15 or more verbatim quotes or complaints from reviews (G2, Trustpilot, app stores, Amazon), Reddit or forums, grouped by pain and ranked by frequency. Exact words only, each with a link and a date
   Pricing	The price ladder in the market, willingness to pay, the most common model, and the price elasticity signals (discounts, churn complaints about price)
   Channels	3 acquisition channels with a cost estimate (CPC, CPM, creator rates, from sourced benchmarks) and real examples of companies using them. Where the ICP gathers (communities, platforms, events). For a mobile app: App Store / Google Play keyword popularity for the target keywords (Apple's floor is 5: all keywords at 5 means search brings almost no one)
   Regulation & platform	Per-country laws, licences, taxes (VAT, OSS), terms-of-service risks, dependency on one platform, with links to the legal text or the platform page. Recent changes (tech, law, platform, price) for why now

3. Update the numbers.
   • Recompute the metrics table (kill skill §4), one by_country set per country, and pick the launch_country with its launch_reason.
   • Re-score all 21 sub-scores with the rubric, the formula and the fixed rules.
   • Recompute confidence. A deep dive should reach high; if it doesn't, say why.
   • The verdict may change. Say what moved it, which sub-scores, and by how many points.
   • Rewrite the premortem, why now and ICP (kill skill §3 and §5) from the new evidence. Every premortem cause links to a kill shot or sub-score of the updated verdict.
4. Go to market (only when the updated score, or the best rescue path, is 50 or more; for a dead idea write "no go-to-market plan: the idea doesn't survive" and skip this step). Build it from the evidence you already have, nothing invented. Save it in report-data.json narrative:
   • personas (2–3): name (a label, e.g. "The solo mandataire"), role, context, pain_quotes (at least 3 verbatim quotes from the customer voice, each {quote, url}), uses_today and pays_today (with price), trigger (the moment they start looking), objections (up to 3), where (up to 5 places to reach them). A persona with fewer than 3 linked quotes is not written.
   • strategies (2–3 ways in): name (beachhead niche, launch country, partner or reseller, service first…), segment, channel, price, cost_first_10 (the cost to win the first 10 customers, sourced or estimate), risks, score_after (re-scored with the same rubric and rules, as an estimate), and recommended: true on exactly one, with why.
   • promotion (3–5 messaging angles): angle, persona (one of the persona names), pain (one of the ranked pains), headline (at most 70 characters), proof and proof_source (a sourced fact that backs the headline), channel, and beat: up to 3 competitor messages it must beat, quoted from their sites ({competitor, message, url}). Plus landing: {headline, subhead, sections (up to 5), cta} for the 7-day test page.
   • branding: positioning {for, category, difference, proof}; names: exactly 5, each checked by search only on the day: domain_com and domain_local (available, taken or unknown), trademark (clear, conflict or unknown, from a search of the launch country's register: USPTO, EUIPO, INPI, UKIPO…) with trademark_source, handles (App Store, X, Instagram, TikTok: what you found), checked (the date). And tone: 3–5 words for the voice, taken from how the personas talk. Never register, buy or reserve anything.
5. Write deepdive.md next to report.md, with these sections:
   1. Updated card.
   2. What changed since the quick verdict (score before → after, the sub-scores that moved).
   3. Market, why now and ICP.
   4. Competitor teardown table.
   5. Customer voice: pains ranked by frequency, with quotes.
   6. Pricing ladder.
   7. Channels with costs.
   8. Country comparison and the launch country.
   9. Risks and premortem.
   10. Updated rescue paths, 7-day test and kill criteria.
   11. Go to market: personas, strategies (the recommended one first), promotion angles with the landing page, and branding (positioning, the 5 names with their checks, tone).
   12. Sources: at least 40 from 15 domains, each with its tier and access date.
6. Update verdict.json and the HTML report.
   • In verdict.json: append the previous evaluation to history, set mode to "deepdive" and evaluated to today, and refresh the metrics, sub-scores, score, rules, confidence, rescue paths and sources.
   • Rewrite report-data.json (kill skill §6). Add customer_voice, the full competitor list with price_low, price_source and traction, the by_country breakdown, and the updated premortem, why_now and icp.
   • Rebuild report.html with python3 "${CLAUDE_PLUGIN_ROOT}/scripts/build-report.py" ideas/<folder> (kill skill §6). An older folder with report-data.js is moved to JSON first with python3 "${CLAUDE_PLUGIN_ROOT}/scripts/migrate.py" --json ideas/<folder>. The HTML then shows the deep-dive version, with the score history.
7. Update the brain: python3 "${CLAUDE_PLUGIN_ROOT}/scripts/build-brain.py" ideas when Bash is available (kill skill §8).
8. Self-check with python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check-idea.py" ideas/<folder> when Bash is available; fix every error before answering.
9. Answer in chat:
   • the updated card;
   • the score change ("42 → 47, because…") and the confidence level;
   • the 3 most surprising findings;
   • the launch country;
   • the recommended way in (strategy), the strongest promotion angle, and the name with the cleanest checks (or "no plan: the idea doesn't survive");
   • the path to report.html (open it in a browser) and to deepdive.md.

Evidence rules are the same as in the kill skill: every number is sourced or marked (estimate) with its formula, unknowns stay unknown, quotes are verbatim. Never publish anything, spend money, or contact anyone.