Analyzing DTC Stores

Produces an investor-grade teardown of a DTC brand from its public URL. Read-only — never writes to the store, sends email, or posts anywhere.

Inputs

• Required: store_url — brand's primary storefront.
• Optional: depth — quick | standard (default) | deep.
• Optional: focus — free-text bias (e.g. "supply chain", "acquisition target fit").
• Optional: --no-save — return the report inline and skip saving.

Process

Step 1 — Intake

1. Parse store_url, depth, focus, --no-save.
2. Derive brand_slug from the domain (e.g. brandname.com → brand-name).
3. Copy assets/report-template.md to a working draft. Do not save to the final path yet.

Step 2 — Recon sweep (scripted)

Run scripts/recon.py --url <store_url> to fetch homepage, robots.txt, sitemap, JSON-LD, and detect platform + apps (Shopify/Klaviyo/Gorgias/Recharge/Triple Whale signatures). Output cached in .cache/recon-[slug].json. Feeds the Tech Stack and Agentic Readiness sections.

Step 3 — Fan-out research (scripted + reasoning)

Load references/sources-playbook.md — required reading, it maps each report section to prescribed sources.

Run every applicable script. Fail gracefully: if a source is unreachable, log [source unavailable] in the report's Sources section and continue — never halt.

Mandatory scripts (all depths):

• scripts/meta_ad_library.py --brand <brand> — active ad count, creative lifespan distribution.
• scripts/importyeti_lookup.py --brand <brand> — supplier + country-of-origin + shipment volume. Mandatory — the #1 skipped source.
• scripts/similarweb_lookup.py --domain <domain> — traffic with the 50K/mo accuracy floor flag.
• scripts/store_leads_lookup.py --domain <domain> — Shopify plan, app stack, revenue bracket.
• scripts/reviews_scan.py --brand <brand> --domain <domain> — Trustpilot + Amazon + YouTube review URLs.
• scripts/reddit_search.py --brand <brand> — Reddit sentiment via OpenAI web search (requires OPENAI_API_KEY). Returns structured JSON: thread_count, sentiment, top_praise_theme, top_complaint_theme. "0 threads" is a valid finding — log it in §10.

Additional for `standard` and `deep`:

• scripts/amazon_bsr.py --brand <brand> — BSR + est. monthly unit sales for top SKUs.

For the reasoning steps (brand story, competitive strategic group, verdict framing) use WebSearch + WebFetch directly.

Step 4 — Unit economics

Read references/unit-economics-benchmarks.md for category COGS / CAC / LTV ranges. Run scripts/unit_econ.py --category <cat> --aov <aov> --cac <cac> to produce a CM1 → CM2 → CM3 waterfall. Every figure must carry stated assumptions and a ~ (est., method: …) prefix.

Step 5 — Draft the report

Fill the 16-section template in order. Principles:

• Prefer "Insufficient public data" over speculation. A 2-line section with the sources checked beats 8 lines of filler.
• Cite inline with [1][2] — the Sources section at the end is the numbered index.
• Every numeric claim is either cited or prefixed `~` with method shown. Non-negotiable.
• Competitors must come from independent discovery, not the brand's own "vs" page. If the brand lists competitors on its site, exclude them until resurfaced via search intent or category overlap.
• Show the CM waterfall, not just gross margin. Apparel especially — returns collapse CM2.
• SimilarWeb under 50K/mo = directional only. Quote a range, not a point estimate; triangulate with ad count + review velocity + headcount.

For depth=deep: after the draft, present three alternative verdict framings (bull, bear, contrarian) and let the user pick before finalising.

Step 6 — Self-check (scripted, mandatory)

Run scripts/validate_report.py --path <draft>. The script fails loud with a remediation list if:

• Any of the 16 sections is missing or empty.
• Any numeric claim is unsourced and not prefixed ~.
• The Sources section is empty or has fewer than 5 entries.
• The Verdict section lacks a numerical "what would change my mind" anchor.
• Any banned hype word appears (revolutionary, disruptive, game-changing, cutting-edge, seamless, unparalleled).
• The ImportYeti check is missing from the Sources section (even a "no records found" entry counts).

Fix every failure and re-run. Do not proceed to Step 7 until the validator exits 0.

Step 7 — Save + return

Run scripts/save_report.py --path <draft> --slug <brand_slug> (honors --no-save). Returns the final path. Then output to chat: final path + a 5-bullet executive summary (≤120 words total).

Step 8 — Closing feedback gate

Ask the user exactly once: "Any corrections to this teardown before I log it?"

Route the response:

• Behavioural feedback (tone, framing, depth preference) → append to references/learnings.md.
• Factual exception (platform quirk, blocked source) → append to references/edge-cases.md.
• "Never again" → add a rule to this SKILL.md.
• Approval with no changes → copy final report to assets/approved-examples/.

Report structure (16 sections, exact order)

1. Research confidence & source map
2. Executive summary (5 bullets, ≤120 words)
3. Brand & positioning
4. Product & value proposition
5. Market & category (TAM/SAM, growth, tailwinds, regulatory)
6. Competitive landscape (5–10 competitors in a table; name the strategic group)
7. Pricing & unit economics (CM1/CM2/CM3 waterfall, CAC, payback — show the math)
8. Supply chain & sourcing (ImportYeti, FDA, LinkedIn, job ads)
9. Channel mix (DTC, Amazon, marketplaces, wholesale, international)
10. Marketing & traffic (Meta Ad Library, Similarweb, creative lifespan, SEO)
11. Customer voice & sentiment (review volumes, themes, NPS proxy)
12. Tech stack
13. Agentic / AI readiness (JSON-LD density, agentic storefront signals, checkout friction)
14. Risks & red flags (incl. manufacturer-direct threat, tariff exposure)
15. Financial summary & valuation (EV/Sales comps, buyer-type implications)
16. Investor verdict (bull case, bear case, numerical "change my mind" anchor)
17. Sources (numbered, with access dates)

Style: crisp, evidence-led, numerate, slightly sceptical. Voice of a senior analyst writing for a partner with 10 minutes. Length: 1,500–3,000 words for standard.

Gotchas

• Symptom: report states "$X ARR" without a source. Cause: triangulation mistaken for fact. Fix: every revenue figure cites a filing/interview/press release, or prefix ~ (est., method: …) and show the math.
• Symptom: competitor list matches the brand's own "vs" page. Cause: used brand marketing as source of truth. Fix: build the competitor set from search intent + category overlap; exclude any competitor named on the brand's own site until independently surfaced.
• Symptom: unit economics cite gross margin only. Cause: stopped at CM1. Fix: always show the CM1 → CM2 → CM3 waterfall with stated assumptions; apparel especially — 24–30% return rates collapse CM2.
• Symptom: SimilarWeb precise visitor count for a small brand. Cause: ignored the 50K/mo accuracy floor. Fix: below 50K/mo, quote a range and label "directional only"; triangulate with ad count, review velocity, LinkedIn headcount.
• Symptom: no ImportYeti entry in Sources. Cause: forgot the mandatory check — the single most commonly skipped DTC source. Fix: Step 3 blocks until importyeti_lookup.py has run; a "no shipment records found" result still counts.
• Symptom: verdict hedges ("interesting brand, execution risk at scale"). Cause: analyst avoided commitment. Fix: verdict must contain a numerical anchor (e.g. "investable below $80M valuation") AND a "what would change my mind" line. Validator fails otherwise.
• Symptom: Meta-reported ROAS cited without qualification. Cause: ignored post-iOS14 attribution collapse. Fix: treat brand-reported Meta ROAS as directional; note the ATT caveat; prefer MER or incrementality-adjusted estimates.
• Symptom: blended CAC cited instead of new-customer CAC. Cause: blended metric flatters because returning-customer spend is in the denominator. Fix: separate the two; new-customer CAC is the scaling metric investors care about.

Rules

• SKILL.md is the routing layer. Domain knowledge lives in references/ — load on demand, don't embed.
• references/sources-playbook.md is required reading at Step 3. Not optional.
• scripts/validate_report.py is a hard gate at Step 6. Not optional.
• ImportYeti check is mandatory even on quick depth — a "no records found" entry is an acceptable result, silence is not.
• Never call this skill "complete" until the validator exits 0.
• Read-only always. Never POST to the store, send email, or publish.
• Banned hype words in output: revolutionary, disruptive, game-changing, cutting-edge, seamless, unparalleled.

<!-- Built with Agent Engineer Master — get your own production-ready skill: www.agentengineermaster.com/skill-engineer -->