growth-factors-mining

name	growth-factors-mining
description	Builds a per-run LIGHT growth-vector database for the Diffmode growth-tactics pipeline by mining public growth case studies fresh, every run, and distilling each into atomic "growth factors" (transferable mechanisms). Clean-room — NEVER reads the proprietary tactics_DB. Outputs growth-factors.json (~20-40 vectors spread across the 6-category scheme — not every category is populated; `conv-` is routinely empty on a demand-gen run) in the schema the synthesis chain + lite-constraints consume. Use as the per-run substitute for the proprietary 576-vector database when generating free growth-tactic ideas.
metadata	{"version":"1.0.0"}

Growth-Factors Mining (per-run LIGHT vector DB)

You build a small, fresh growth-mechanism database from public case studies, distilled using the "mechanism over tactic" method. This is the free pipeline's substitute for the proprietary 576-vector database: a deliberately weaker, clean-room asset that gives the synthesis chain real vectors to combine without shipping any proprietary IP.

⚠️ Clean-room rule (moat-critical — non-negotiable)

You MUST build this only from freshly researched public sources. You MUST NOT read, open, glob, or grep anything under tactics_DB/ (the proprietary vector DB, its intelligence layer, and anti-vector tracking — and any script that reads them). No content here may be traceable to that database. The value you ship is the method; the DB it produces is intentionally lighter than the paid one. If any input path points into tactics_DB/, refuse it and note it in your summary.

Inputs & Output

The invoker provides (do not hardcode absolute paths):

• INPUT — founder context (required): WS/01-diagnostics/founder-input.md. Read FIRST — use the product's business model, industry, stage, channels, and audience to bias your case-study search toward relevant growth stories (a bootstrapped B2B SaaS should mine indie SaaS / community-led / content / PLG case studies, not enterprise ad-spend stories).
• INPUT — competitive context (optional but recommended): WS/02-enrichment/competitors-analysis.md and WS/02-enrichment/acquisition-tactics.md — to seed searches around the channels/tactics live in this founder's space and the adjacent industries worth borrowing from.
• WEB RESEARCH (required capability): your web-research backend — search plus page retrieval. This is the ONLY source of vectors.
• OUTPUT: write WS/03-think-tanks/demand-generation/growth-factors.json.

Caching & bounded research (cost control — surface this tradeoff)

This stage is deliberately "fresh per run," which is slower / pricier / less deterministic than a static asset. Mitigate:

• Cache: if growth-factors.json already exists at the output path AND the brief does NOT set remine: true, do NOT re-research. Read the existing file, validate it against the schema + counts below, and return {status:"ok", ...,"summary":"reused cached growth-factors.json (N vectors)"}. Re-mine only when the orchestrator brief sets remine: true (a Start-fresh relaunch) or the file is missing/invalid.
• Resume-partial (socket-death recovery — do NOT re-pay for deep research): if the brief includes resume_partial: true (the orchestrator sets this only when re-spawning after a mid-mine death) AND a partial growth-factors.json exists that PARSES but is short of target (e.g. < 20 vectors, or otherwise incomplete), read it, KEEP every already-distilled vector verbatim, and mine ONLY the remainder needed to reach the target count + category spread. Continue each prefix's sequential numbering from where the partial file left off; do NOT re-run the deep-research passes that produced the vectors already on disk — that duplication (≈8 deep research calls) is exactly what this flag exists to avoid. resume_partial is never combined with remine: true (which forces a full fresh re-mine); if both somehow appear, remine wins and you re-research from scratch.
• Bound breadth + cap deep research (the run's biggest cost lever): review 12-20 public case studies using at most ~1-2 deep-research passes (iterate search + page retrieval, or a single deep multi-source call if your backend has one) — seed them from the founder context for the initial case-study landscape, then gather the remaining case studies + their specific metrics with cheaper plain-search calls. Do NOT open-ended crawl. This stage's deep-research calls were the single biggest cost driver in the field (~85% of a run's research spend; a socket-death respawn used to duplicate them), so keep them scarce — search-first. Stop when you have enough distinct mechanisms to hit the target count.

Method — adapt the proven extraction methodology

Apply the proven "mechanism over tactic" extraction method (inlined below). For each case study:

1. Mechanism, not tactic. Capture WHY it worked at a first-principles level ("high-concept analogy bypasses explanation friction"), NOT what they did ("posted on LinkedIn"). One case study yields 1-3 atomic vectors.
2. Transferability test. Would this work in a completely different industry? Give 2-3 cross-industry examples proving it transfers. Drop "Low transferability" / very context-specific findings.
3. Dedup. Before adding a vector, check it isn't the same mechanism as one already in your list with different words. Merge duplicates; keep the cleaner statement.
4. No generic advice. Reject "be consistent", "post regularly", "talk to customers" — those aren't vectors.
5. Demand-gen lean. Prefer lever- / resource- / struct- (acquisition/distribution mechanisms). psych- and pos- are allowed when the mechanism drives acquisition (reciprocity → partnership access, exclusivity → community growth, authority signaling → outreach acceptance). conv- only for the rare acquisition-adjacent conversion mechanism. This is a demand-gen pipeline.

Categories & ID format (NOT secret — reused so synthesis runs unchanged)

The 6 categories and the {prefix}-NNN-slug ID format are public conventions. Reuse them so the ported synthesis prompts consume your output unchanged. Numbering is local to this run — number sequentially per prefix starting at 001 based on the order you mine them. Any resemblance to proprietary IDs is incidental; you derive these independently.

Category	Prefix	Mechanism is about…
Structural Arbitrage	struct-	Timing, platform/market gaps, competitive positioning windows
Leverage Mechanisms	lever-	Compounding, amplification, flywheels, viral loops, network effects
Resource Optimization	resource-	Efficiency, validation, risk reduction, lean/bootstrapped execution
Psychological Mechanisms	psych-	Cognitive biases, trust, urgency, social proof (acquisition-side)
Positioning Dynamics	pos-	Differentiation, framing, anchoring, contrarian positioning
Conversion Architecture	conv-	Funnel/offer mechanics that drive acquisition (rare here)

Target output

• 20-40 vectors total. Aim for spread: a healthy run has the majority in struct-/lever-/resource-, with a few psych-/pos-. No single prefix should exceed ~60% of the vectors. If you can't responsibly reach 20 distinct, transferable mechanisms from public sources, write what you have (≥15) and note the shortfall.
• Not every category will be populated, and that is correct. The spread target is an upper bound on concentration, not a requirement that all six prefixes be non-empty. conv- is routinely 0 on a demand-gen run (see the demand-gen lean above), and pos- is often low single digits. Emit all six keys in category_counts with their real values — including 0 — and never invent a vector to fill a category.
• Each vector carries the schema below, with real evidence + a source URL (this is how the output proves it's clean-room and not invented).

Output schema (write EXACTLY this JSON shape)

{
  "metadata": {
    "generated_for": "<workspace slug / product>",
    "generated_date": "<YYYY-MM-DD>",
    "method": "clean-room per-run mining from public case studies (mechanism-over-tactic)",
    "source_note": "Diffmode growth-tactics LIGHT DB. Built fresh from public case studies. NOT the proprietary 576-vector database.",
    "case_studies_reviewed": <int>,
    "total_vectors": <int>,
    "category_counts": { "struct-": 0, "lever-": 0, "resource-": 0, "psych-": 0, "pos-": 0, "conv-": 0 }
  },
   
    
       
       
       
       
       
       
       
        
        
        
      
       
       
       
    
  

Field rules (from the extraction methodology): mechanism = 1-2 sentences, transferable, not case-specific; transferability ∈ {High, Medium, Low} (avoid Low); saturation_risk ∈ {Emerging, Mature, Oversaturated}; examples = 2-3 in DIFFERENT industries than the source; evidence quotes/paraphrases the actual case study (numbers when available — never fabricate); source_url is a real, reachable URL; time_to_signal_weeks optional integer.

Procedure

1. Read founder context (+ competitive context if provided). Derive 4-6 search themes (business model, primary channels, industry, adjacent industries to borrow from).
2. Check the cache / resume-partial (see above). If a valid full file exists and the brief does not set remine: true, reuse and return. If the brief sets resume_partial: true and a parseable but short partial file exists (and no remine), load it, keep its vectors, and mine only the remainder — skip the deep-research passes for what's already there.
3. Deep research pass (search-first, ≤~1-2 deep calls): run a bounded set of web-research calls — at most ~1-2 deep multi-source passes for the initial landscape, then cheaper plain-search calls — on growth case studies across those themes + 2-3 deliberately different industries (for transferable mechanisms). Capture source URLs + the specific result/metric for each story. Guerrilla search seeds: alongside the founder-derived themes, include at least one search pass using unconventional/guerrilla angles — e.g. "ambush marketing case study," "counter-cyclical launch timing," "secret menu / exclusive offer growth," "community infiltration marketing," "mystery benefactor / anonymous giveaway," "reverse review / customer-as-hero marketing," "hyperlocal guerrilla tactic," "partnership judo startup." These pull in case studies the default "growth case study" query misses (physical-world, event-based, psychological, and partnership mechanisms).
4. Distill each case study → 1-3 atomic vectors using the method above. Assign category
   • a local sequential {prefix}-NNN-slug id. Write mechanism, transferability, saturation_risk, 2-3 cross-industry examples, evidence, source_url.
5. Breadth check. Map every distilled vector to a mechanism type: content/SEO · partnership/alliance · timing/counter-cyclical · pricing/offer · community/tribe · outbound/direct · event/experiential · platform/technical · psychological/behavioral · structural/regulatory. If any type that the founder's industry could plausibly use has ZERO vectors, do one more targeted search for case studies in that type before proceeding. This is a check, not a constraint — new vectors still must pass the transferability test and the mechanism-over-tactic rule.
6. Dedup + balance to 20-40 vectors with category spread (no prefix > ~60%).
7. Compute metadata (counts, category_counts, case_studies_reviewed) and write valid JSON to the output path. Validate it parses (-clean).

Validation checklist (self-check before returning)

• Output is valid JSON in the exact schema above; total_vectors matches vectors length; category_counts sums to total_vectors.
• 20-40 vectors (or ≥15 with a noted shortfall); no single prefix > ~60%.
• Breadth check ran: vectors span ≥5 distinct mechanism types (content, partnership, timing, pricing, community, outbound, event, platform, psychological, structural); any plausible type with zero vectors triggered a follow-up search.
• Every vector is a MECHANISM (WHY), not a surface tactic (WHAT).
• Every vector has 2-3 cross-industry examples, real evidence, and a real source_url. No fabricated sources or metrics.
• Demand-gen lean (majority struct-/lever-/resource-); psych-/pos- only for acquisition-side mechanisms.
• CLEAN-ROOM confirmed: nothing was read from tactics_DB/; nothing is traceable to it.