A thin orchestration skill for commercial/market research. It triages your topic, finds the right specialized data source (and helps you install it), then hands the heavy lifting to the research harness you already have — instead of reinventing it.
market-intel is built on one principle — root-cause design, not incremental patching. When something is wrong, we change the assumption underneath it, not the symptom on top. That single idea produced every decision here: browser-automation was promoted from footnote to a first-class route (not "add a few free tools"); this is a thin delegation layer (not "another deep-research"); updates run through a deterministic gate that can only let the matrix improve (not "set a reminder to refresh"). The philosophy outranks any individual feature — every future change must pass one test: does it fix the framing, or just patch a symptom?
📜 Read the full design philosophy → PHILOSOPHY.md (6 principles, each with the patch-vs-root contrast and the real decision in this repo that it produced).
Claude Code already has a deep-research harness (fan-out → fetch → verify → synthesize) and a research-lit skill. Those are great for general web and academic research. They fall short the moment your question needs a specialized commercial source behind an information barrier — real X/Twitter data, Amazon price history, on-chain feeds, SEO metrics, social sentiment, B2B lead data.
market-intel is the thin layer that fills exactly that gap. It does only three things nothing else does, and delegates everything else:
- Triage — map a commercial topic to 1–N of 12 data domains.
- Detect + guide install — check which specialized MCP sources are actually connected (via
claude mcp list, not unreliable tool-name guessing), and if a key source is missing, hand you the exactclaude mcp addcommand. - Quality guardrails — citation verification, source tiers, multi-source corroboration, mandatory disconfirmation, explicit gaps.
The actual fan-out, fetching, adversarial verification, and citation synthesis are delegated to deep-research / research-lit. No reinvented engine, no trigger fights.
/plugin install github:DaizeDong/market-intel
Or clone manually:
git clone https://github.com/DaizeDong/market-intel.git ~/.claude/plugins/market-intelIt auto-activates on phrases like 市场调研, competitor analysis, research this market, find arbitrage opportunities, X/Twitter sentiment, SEO intel, product trends. For single-fact lookups or general web reports it deliberately steps aside (use plain search / deep-research); for academic literature it defers to research-lit.
You say:
research the competitive landscape and X sentiment around <product>, then find any arbitrage angle
What runs:
- Triage → maps to
x-twitter,trends-discovery,ecommerce-arbitrage; picks a depth budget with hard caps (no runaway fan-out). - Detect → runs
claude mcp list, sees you have none of the X/ecommerce MCPs connected, notes it. - Guide install (non-blocking) → "This depends on real X data. Install twitterapi.io:
claude mcp add -s user ...— note it only works after a session reconnect. For now I'll use web fallback and flag the gap." - Delegate → fans out subagents / invokes
deep-research, each returning a structured evidence unit (claim · source · quote · tier · date · confidence), not raw page dumps. - Guardrails → independent verifier re-fetches each cited URL; decision-grade claims need ≥2 independent sources; a dedicated reverse-search subagent hunts risks/failures.
- Report → snapshot-dated, tier-tagged, with a disagreement matrix, a mandatory Risks & counter-evidence section, and an explicit "configure source X for deeper data" gap list.
The knowledge asset. Each domain shard names the best tool, its barrier route, how to detect it, and what to install. Thin index → load only the domain(s) you need.
| Domain | Top pick (barrier route) |
|---|---|
| x-twitter | twitterapi.io ② resale |
| reddit-community | HN MCP ① free · Reddit API ① |
| web-scraping | Tavily/Exa + Firecrawl + Bright Data |
| ecommerce-arbitrage | Keepa ① official |
| finance-markets | SEC EDGAR + FRED ① free |
| crypto-defi | CoinGecko ① + ccxt |
| seo-keywords | GSC ① free + DataForSEO ② |
| social-publishing | Buffer ① · Postiz OSS |
| content-cms | Sanity/WordPress MCP ① |
| leadgen-crm | Apollo.io ① + Hunter ① |
| trends-discovery | GDELT + Product Hunt MCP ① free |
| ready-skills | coreyhaines31/marketingskills |
| browser-automation | playwright MCP + browser-use / crawl4ai ④ |
Barrier routes: ① official API (compliant, often paid) · ② resale API (provider absorbs the barrier, cheap, gray-area) · ③ self-host scrape (reverse-engineered API, free, accounts+proxies, ban risk) · ④ browser automation / act-like-human — real logged-in browser (playwright MCP + free OSS repos). First-class, not a footnote: often returns richer data (rendered/logged-in view, fields APIs hide) at zero API cost. The skill prefers route ④ over paid APIs when it fits, reaching for ①/② only for history it can't backfill (e.g. Keepa), scale reliability, or compliance.
Exact install commands and prices live in pricing-install.md, each line last_verified-stamped — verify against the official site before quoting.
Hard rules applied during synthesis (see SKILL.md):
- Citation verification gate — an independent verifier re-fetches every cited URL and confirms the page contains the value (verbatim quote). Dead links dropped; quote-less numbers demoted to "unverified."
- ≥2 independent sources for decision-grade claims; each tagged confidence high/medium/low.
- Source tiers L1 first-party → L5 fallback/inference; vendor self-claims can't be sole support.
- No silent degradation — falling back from a barrier source to web is flagged in-line.
- Timestamp volatile data — every price/policy carries fetched + published dates.
- Disconfirmation mandate — a reverse-search subagent hunts scam/failure/risk; arbitrage gets an explicit execution-friction section.
- Surface conflicts, don't average them; failures become explicit coverage gaps.
The matrix decays — APIs go paid, tools get acquired, prices move. The refresh protocol re-sweeps each domain (one subagent per domain → structured diff → incremental shard edits → CHANGELOG.md + version bump). Default cadence quarterly; monthly for fast-moving domains (x-twitter, web-scraping, social-publishing, crypto-defi). Trigger manually with 刷新工具库 / refresh the market-intel source matrix, or wire a scheduled headless run (see ROADMAP).
This skill is the product of a 12-subagent tool survey followed by a 5-subagent adversarial design review. The review killed the original "build another full deep-research" plan (it would have been a clone with a trigger conflict), proved that claude mcp add doesn't take effect until a session reconnect, and forced in the citation-verification gate, source tiers, and disconfirmation mandate. See ROADMAP.md for what's next.