nerissa-competitive-research

# nerissa-competitive-research · Disciplined Competitor Research SOP A field-tested SOP for competitive research that refuses lazy AI shortcuts. Built from real mistakes — sourcing from secondhand blogs, inheriting unverified DD judgments, confusing funding size with relevance, treating "in negotiation" as "signed". Every rule in this skill traces back to a specific failure mode this skill prevents. ## When to Trigger **✅ Trigger when**: - User explicitly says: "research X company", "竞品调研", "调研 X 家", "do competitive analysis", "stress-test this competitor" - Adding any new competitor to an existing comparison table or matrix - Validating a competitor claim before publishing (pitch deck / report / website) - User questions a previous competitive judgment ("are you sure about X?") **❌ Do NOT trigger for**: - Casual market discussion without a specific company - User explicitly asks for a brief one-line answer - General industry trend questions ## Core Principles (Earned from Real Failures) | # | Principle | What It Prevents | |---|---|---| | 1 | **Every "signed / acquired / valuation / customer count" claim needs verifiable source** | Repeating "$500k seed" from a DD report when reality is "$0 bootstrapped" (one secondhand blog was wrong) | | 2 | **Open the official website. Don't rely on aggregator blogs** | Treating an anti-scraping 403 as "no website exists" and skipping the company entirely | | 3 | **Inheriting historical judgment requires explicit tagging** | Recommending a competitor on old DD without re-evaluating that it's actually a packaging layer over commodity tools | | 4 | **Decompose the essence. Don't just describe what it does** | Saying "AI full-stack builder" without saying "essentially a packaging layer over UI-gen + BaaS + edge cloud" | | 5 | **Funding size ≠ relevance** | Treating $400M-ARR companies as learning targets when their customer layer doesn't match yours at all | | 6 | **Customer overlap requires verification, not intuition** | Claiming "their customers don't overlap with ours" when actual overlap is 30-50% | | 7 | **Separate dual tracks: fundraising-reference vs business-learning** | Killing a $0-funded peer because "our fundraising goal is $10M+" — confusing two different uses | | 8 | **Don't say "X has signed Y customers" without checking the source-of-truth database** | Reporting "we signed 3 customers" in 5 reports when source-of-truth showed 0 signed, 2 in quote, 9 in conversation | | 9 | **Tag every claim with Evidence Level (L1A-L5). Don't mix L5 self-reports with L1B verified data** | Treating "iSales 月入百万 (founder self-report)" as same credibility as "Nectar $40.6M (3-source verified)" — leads to fragile strategy built on un-verified claims | ## The 8-Step SOP (40-75 minutes per company) ### Step 1 · Open the official website (30 seconds) - Open the main domain - Screenshot the hero - If site doesn't load / returns 403 → mark this fact, don't assume "no website" ### Step 2 · Find funding data (5 minutes) **Source priority** (most to least trustworthy): 1. Crunchbase company page (primary) 2. Official funding PR / TechCrunch / Bloomberg 3. PitchBook (paid, summary visible) 4. LATKA (CEO self-reported, best for bootstrapped SaaS) 5. Secondhand aggregator blogs (use cautiously, require ≥ 2 independent sources to cross-verify) **Required fields**: cumulative funding / current valuation / latest round (amount + date + lead) / ARR (if disclosed) ### Step 3 · Decompose the essence (10 minutes) Ask 5 questions: - What is it **essentially**? (technology / packaging layer / sales motion / brand / network effect) - If it disappeared, what would happen? (many substitutes / high switching cost) - What type of moat? - Where does it sit on the 5-dimension category map? (service-delivery / customer-tier / pipeline-depth / vertical / business-model) - In the current funding wave, what did it leverage? (technology / packaging / timing / capital) ### Step 4 · Verify customer overlap (10 minutes) - Check customer logo wall on the official site - Read 1-3 case studies to identify actual customer identity / size / pain point - Search Twitter / Reddit / G2 / ProductHunt for real user feedback - Cross-reference against your own current customers to assess overlap ### Step 4.5 ⭐ NEW · Tag every claim with Evidence Level (5 minutes) Before scoring, classify every piece of evidence you collected into one of 5 levels: | Level | Source Type | Example | |---|---|---| | **L1A** | Audited financials / SEC / IPO prospectus + ≥ 2 sources | "Snowflake ARR $2.8B per 10-K filing" | | **L1B** | Tier-1 financial press cross-verified ≥ 3 sources | "Nectar $40.6M cumulative per TechCrunch + Menlo + Yahoo" | | **L2** | Official site + company PR + LinkedIn page ≥ 2 sources | "Nectar serves DTC brands per nectarsocial.com logo wall + PR" | | **L3** | Single trusted vertical media / G2 reviews | "百型智能国家备案 per 新浪财经 single-source — pending cross-check" | | **L4** | Secondary aggregators / VC blog / founder interviews | "iSales founder background per Tencent News interview" | | **L5** ⚠️ | Self-report / Claude inference / prediction | "iSales 月入百万 (founder self-report ⚠️)" / "I estimate ARR $5-20M" | **Mandatory output**: Every assertion in steps 1-4 must carry a `[L1A/L1B/L2/L3/L4/L5]` tag. See `references/evidence-level-pyramid.md` for full rules and examples. ### Step 5 · Score on 6 dimensions (5 minutes) Score 0-2 on each, sum to 12. Only 8+ qualifies for Top. **⚠️ NEW**: Each dimension's score must also include its supporting **evidence level**. Final score reported as both **nominal** and **discounted** (L5-backed dimensions = 0 in discounted). | Dimension | 0 points | 1 point | 2 points | |---|---|---|---| | Customer overlap | Completely mismatched | Partial overlap | Highly overlapping | | Business model borrowability | Path entirely different | Some shared mechanics | Directly copyable | | Capital/team scale proximity | $100M+ / 100+ people | AI-50M / 30-80 people | <$10M / 2-20 people or bootstrapped | | Recent activity | Dead / 18 months silent | 12 months product updates no funding | 6 months funding / ARR disclosed | | Real cash flow | All funding-burn | ARR > $1M but funding-supported | Bootstrapped or ARR/funding > 1 | | Pipeline relevance | Unrelated | Part of pipeline relevant | Full pipeline (Concierge) directly relevant | **Score table format** (new mandatory column): | Dimension | Score | Evidence Level | Rationale | |---|:---:|:---:|---| | Customer overlap | 2 | L2 | Official site logo wall + case study | | Cash flow | 2 | L5 ⚠️ | Founder self-report on Tencent — UN-VERIFIED | | **Nominal total** | **11/12** | — | — | | **Discounted total** | **9/12** | — | L5 cash flow dimension contributes 0 | ### Step 6 · Produce 3 sentences (15 minutes) - **One-sentence "what it is"** - **One-sentence "what to copy"** (the single most important thing for you) - **3 must-see pages** (let user visit themselves, not hear secondhand) ### Step 7 · Enter into structured database (5 minutes) Use the 24-field schema (see `references/feishu-base-schema.md` for Feishu Base template). **NEW mandatory fields** (added 2026-06-07): - `evidence_level_overall` (L1A/L1B/L2/L3/L4/L5) - `funding_evidence_level` - `customer_evidence_level` - `arr_evidence_level` - `last_evidence_review` (date) - `pending_verification` (multi-line text — list every L5 ⚠️ claim and the action to upgrade it) ### Step 8 ⭐ NEW · Run the Evidence Level Pre-Publish Checklist (5 minutes) Before any report is shared with the team / investors, run through `references/evidence-level-checklist.md`. The report goes back to revision unless all boxes are ✅. **Critical floor checks**: - No strategic conclusion built solely on L4-L5 evidence - All L5 inferences quarantined to "💭 Inference Section" - All "X will happen in N months" predictions rewritten as hypothetical scenarios - "Pending Verification" section enumerates every L5 ⚠️ and its upgrade path **Sanity test**: If I removed all L5 ⚠️ content from this report, would my conclusions still hold? If no, the report is **draft-only — do not share**. ## Dual-Track Structure (Don't Mix) | Track | Purpose | Selection criterion | |---|---|---| | **A · Fundraising reference** | For pitch decks · cite $10M+ similar-direction companies | Look at funding scale only | | **B · Business learning target** | For product/team decisions | Ignore funding · only look at path similarity | Mixing them causes errors (e.g. killing a $0 bootstrapped 7-person $770k ARR company because "our funding goal is $10M+" — but it's track B, not track A). ## Required Output: 15 Mandatory Fields Every researched company must populate these in your tracking database: | Field | Required? | Must verify? | |---|---|---| | Company name | ✅ | - | | Website URL | ✅ | Must have opened | | Founded year | ✅ | Crunchbase / official | | Headquarters | ✅ | - | | Founders | ✅ | LinkedIn / company About | | Team size | ✅ | LinkedIn / Crunchbase | | Cumulative funding | ✅ | **≥ 2 independent sources** | | Current valuation | ⚠️ if available | Official PR only | | Latest round | ✅ | TechCrunch / company PR | | Lead investor | ✅ | Same | | ARR | ⚠️ if available | LATKA / official / prospectus | | Business model | ✅ | Must have opened pricing page | | One-line "what it is" | ✅ | You write | | One-line "what to copy" | ✅ | You write | | 3 must-see pages | ✅ | You pick | | What to learn for your context | ✅ | 3-5 bullets | | Known customers | ✅ | Customer wall on site | | Time to study | ✅ | For user planning | | Overall judgment | Left for user | 1 of 4: copy directly / partial / enough / not applicable | | Data sources | ✅ | ≥ 2 URLs | ## Anti-Lazy Rules (Reject These Behaviors) | ❌ Don't | ✅ Do | |---|---| | Infer "signed / acquired" from memory | Query the source-of-truth database | | Inherit historical DD without re-validating | Open the site again every time | | Recommend based on funding numbers alone | Score 6 dimensions, only 8+ recommended | | Only describe "what it does" | Always ask "what is the essence — packaging / true tech?" | | Estimate customer overlap by gut | Check site cases + Twitter / G2 reviews | | Mix Track A and Track B in recommendations | Separate them, label which track each is in | | Force "vertical X" branding from one sample | Need 3+ samples before generalizing to a vertical | | Mention "anti-competitor-X" framing | Don't free-advertise your competitors | ## Triggers for New Research | Trigger | Should you research? | |---|---| | Industry report mentions new company | 🟡 Depends | | Investor / customer / peer mentions a company | ✅ Immediately | | TechCrunch / 36Kr reports new funding in your space | ✅ Immediately | | Conference brings in a new potential customer | ✅ Research the customer (not as competitor) | | A giant (major e-commerce / social / cloud platform) enters similar space | ✅ Immediately + high priority | | Existing competitor gets new funding / launches / collapses | ✅ Upgrade the existing card | | Team member says "I think X company is impressive" | ⚠️ First ask 6-dimension preview · only 8+ goes to formal research | ## After Research: 3 Required Sync Points Every completed research must update: 1. **Tracking database** (one new row, all fields filled) 2. **Master report doc** (update master + detailed narrative, monthly version bump) 3. **Memory file / project notes** (update "N companies in database" counter) ## How to Use This Skill When invoked, Claude will: 1. Confirm the company name + which use case (Track A fundraising reference or Track B business learning target) 2. Execute the 7-step SOP in order 3. Score on 6 dimensions 4. Produce the 3-sentence output 5. Format the 15 mandatory fields for database entry 6. Flag any anti-lazy rule violations encountered ## Examples Inside This Skill See `examples/` for 5 worked examples (sanitized) from real research sessions: - `01-bootstrapped-saas-found-via-LATKA.md` — Bootstrapped SaaS case (corrected a $500k-seed myth via LATKA cross-check) - `02-funding-wave-packaging-layer.md` — Funding-wave AI builder case (decomposed essence to UI-gen + BaaS + edge cloud packaging) - `03-china-direct-competitor.md` — Direct local-market competitor found via tech-media news, multi-source verified - `04-incumbent-warning.md` — Incumbent giant entry case (treated as warning, not deep research target) - `05-fundraising-reference-vs-learning-target.md` — Enterprise giant vs bootstrapped peer (dual-track example) ## Linked Resources - `references/feishu-base-schema.md` — Feishu Base table-creation script (24 fields, ready to bash) - `references/six-dimension-scoring-worksheet.md` — Print-and-fill scoring template - `references/anti-lazy-checklist.md` — Pre-publish checklist (don't publish until all green) - ⭐ `references/evidence-level-pyramid.md` — 5-level evidence pyramid (L1A → L5) adapted from EBM; mandatory tagging rules for every claim - ⭐ `references/evidence-level-checklist.md` — Pre-publish checklist for evidence discipline; all boxes must be ✅ before sharing ## A Note on Provenance This skill was extracted from a real growth-stage startup's competitive research practice. Every principle in the "Core Principles" table corresponds to a documented failure that was caught and corrected. The 6-dimension scoring system was empirically derived from comparing 12 sample companies and noting which dimensions actually predicted "useful to learn from" vs "irrelevant to learn from." ### 2026-06-07 Update · Evidence-Based Discipline The evidence-level pyramid (Principle 9, Step 4.5, Step 8) was added after Claude (the AI agent) was caught mixing L5 founder self-reports (e.g. "iSales monthly revenue ¥1M") with L1B verified funding data in strategy conclusions, then using L5-backed scoring to recommend "Track A fundraising reference" status. nerissa observed that this propagated un-verified assumptions into all downstream decisions (PRD direction, pitch deck, differentiation arguments). The 5-level evidence pyramid was directly transplanted from evidence-based medicine (Oxford CEBM), where the same anti-pattern (treating expert opinion same as RCT) caused decades of clinical errors before EBM systematized it. The behavioral rule: **never let an L5-backed scoring item win the strategy conversation alone.**