Amazon Business Research

Amazon Business Research — Browser Edition (Master Orchestrator)

What this is

A Claude Skills package that lets Claude run the full Amazon product-opportunity research workflow implemented in the separate Amazon-products engine (/home/zaid/Desktop/Amazon-products in this environment, or wherever the user's checkout lives — locate it before starting; see Prerequisite below), while adding real research capability the engine's own automated providers cannot: Claude Browser web/Reddit research and a Google Trends human-in-the-loop workflow.

This package does not reimplement the engine's business logic. Every gate, threshold, state transition, and economics/ranking calculation is enforced by the engine's own tested Python code (568 passing tests as of this build). This Skill's job — and every specialized Skill under skills/ — is to drive that engine's hunt CLI correctly, interpret its output honestly, and add research breadth the engine cannot get on its own. Read references/core-rules.md before doing anything else; its five rules are non-negotiable.

Prerequisite

This Skill requires shell (Bash) access to a checkout of the Amazon-products engine with its Python virtualenv set up (.venv, dependencies installed) and hunt runnable (hunt --help or python -m engine.cli --help from the repo root with .venv activated). If that's not available in the current environment, say so explicitly and stop — do not attempt to answer research questions from general knowledge instead. This Skill has no value without the engine underneath it.

Canonical flow

DISCOVER CONCEPT
    ↓
DEMAND INTELLIGENCE (skills/demand-intelligence) ── Google Trends human-in-the-loop
    ↓  (advisory gate only — never blocks; see demand-intelligence.md SVC-08)
AMAZON DISCOVERY (skills/amazon-discovery) ── hunt discover
    ↓  (relevance + cheap filters + staged budgets already applied by the CLI)
AMAZON VALIDATION (skills/amazon-validation) ── live snapshot interpretation
    ↓  (reaches ECONOMICS_CHECK state)
COMPETITION ANALYSIS (skills/competition-analysis) ── hunt deepen-start
    ↓  (reaches COMPETITOR_ANALYSIS state)
VOC (skills/reddit-voc + browser-research, or automated voc-prepare/submit)
    ↓  (reaches VOC_ANALYSIS state)
DIFFERENTIATION (skills/differentiation)
    ↓  (reaches DIFFERENTIATION_ANALYSIS state)
RED TEAM (skills/red-team)
    ↓  (reaches RED_TEAM → HUMAN_REVIEW state)
ECONOMICS (skills/economics) ── can run any time ECONOMICS_CHECK+ is reached
SUPPLIER RESEARCH (skills/supplier-research) ── can run any time, needed before QUALIFIED confidence
IP RISK (skills/ip-risk) ── can run any time, needed before QUALIFIED confidence
    ↓
FINAL DECISION (skills/final-decision) ── hunt rank, hunt show, then hand off to the human
    ↓
hunt decide <opportunity_id> approve|reject   ← HUMAN ONLY, never automated

Do not skip a stage's precondition. references/cli-command-reference.md has the exact required-state table; every -prepare/-submit/deepen-start call raises a clear error rather than silently coercing if the precondition isn't met — if that happens, go back and complete the missing stage, don't route around it.

How to run a research cycle

1. Establish the concept. If the user gave a specific niche/query, use it. If they said something like "find me a product worth building a business around," do not immediately start searching — ask what domain/category interests them, or propose 2-3 concrete candidate niches grounded in whatever context you have, and confirm before spending any research budget. Use skills/demand-intelligence/SKILL.md to check whether a concept already has a recorded demand signal (hunt demand-show) before requesting new Google Trends data.

2. Cheap gates first, always. Run demand intelligence and hunt discover (which already applies relevance classification and cheap eligibility filters deterministically) before any expensive step. See references/time-budget-controller.md. Never run browser-brief, deepen-start's competitor snapshots, or supplier/IP research on a candidate that hasn't survived the cheap stages.

3. Advance survivors stage by stage, using each specialized Skill in skills/ in the order shown above. Check state via hunt show <opportunity_id> whenever you're unsure what stage a candidate is at.

4. Track multiple candidates in parallel where it's cheap to do so — discovery and relevance filtering naturally produce several candidates from one run; don't force single-candidate tunnel vision. But do not run expensive stages (browser research, supplier/IP, Google Trends requests) on more candidates than the budget discipline in references/time-budget-controller.md allows.

5. Stop and reject early when a candidate fails a hard gate or a red-team finding is decisive. Don't keep researching a candidate the engine has already told you is REJECTED.

6. Never reach a final verdict yourself. When a candidate reaches HUMAN_REVIEW (or you want an interim check via hunt rank), use skills/final-decision/SKILL.md to compile the full evidence report and present it — then stop and let the human run hunt decide. Do not draft the decision as if it were already made.

What must never happen (see references/core-rules.md for the full contract)

• A hard gate never fires on UNKNOWN/SUSPECTED_*/PENDING_* — only literal CONFIRMED_* values, and only a human sets those (via hunt supplier-set) or the engine's own lineage heuristics (capped at SUSPECTED_CONTAMINATION, never CONFIRMED).
• QUALIFIED is reached only by the human running hunt decide approve. This Skill package never runs that command autonomously and never implies a recommendation is a final decision.
• --mode real failures are never silently retried with --mode mock.
• Economics numbers are never computed by Claude — only reported from hunt economics-set/economics-show output.
• Evidence from Claude Browser research is recorded via hunt browser-evidence-add, capped at ESTIMATE, never presented as FACT.
• Google Trends' 0-100 index is never described as absolute search volume or sales.

Reference index

Skill index

skills/demand-intelligence, skills/amazon-discovery, skills/amazon-validation, skills/competition-analysis, skills/browser-research, skills/reddit-voc, skills/differentiation, skills/red-team, skills/economics, skills/supplier-research, skills/ip-risk, skills/final-decision — each is independently invocable; this Master Skill sequences them.

See examples/example-research-workflow.md for a full worked run.