Market-Research Skill Review: 244K Stars Worth of Process, or Just a Better Template?

I went looking for a market research skill because I was tired of Claude Code producing gorgeous, useless research dumps. The kind where you ask for competitor analysis and get a five-paragraph essay on each competitor with no numbers, no sources, and zero advice. The market-research skill from the ECC suite by affaan-m is currently peaking on SkillsMP with 244,512 stars and 1,475 gained in the last week. That's not a quiet community vote — that's a thunderclap. So I sat down and read the SKILL.md, ran a few prompts through it mentally, and tried to figure out whether this thing actually changes agent behavior, or if it's just well-formatted instructions.

Short version: it changes behavior. The framing matters more than the checklist.

What this skill does, in plain language

market-research is a behavioral instruction set for an agent. It doesn't ship data, doesn't call external APIs, and doesn't scrape anything. What it does is impose a discipline on how the agent does research: cite everything, surface counter-evidence, separate fact from inference from recommendation, and end with a structure that maps to a real decision.

The opening line of the SKILL.md is the thesis: "产出支持决策的研究，而非研究表演" — research that supports decisions, not research theater. That's the whole pitch in eight characters. If you've ever gotten a long, confident, source-less market analysis from an LLM, you know exactly what "research theater" means.

It activates when you're doing any of: market sizing, competitive analysis, investor/fund diligence, TAM/SAM/SOM estimation, or tech/vendor scanning. The output format it enforces is: exec summary, key findings, implications, risks, recommendations, sources. That structure is nothing new — but the quality gates attached to it are what make it useful.

Why this matters

The default failure mode of LLM research is confidence without grounding. An agent will happily invent a TAM figure, attribute it to "Gartner 2024," and tell you your idea is a $50B opportunity. This skill doesn't fix the hallucination problem at the source level, but it does force the agent to label every claim as either sourced or estimated, and to call out stale data explicitly. That's a meaningful behavior change.

The other thing it does well is the "find the bear case" step. Point 3 of the research standards says: "包含反面证据和不利情况" — include counter-evidence and adverse cases. Most agent outputs are cheerleaders. This skill explicitly instructs the model to play devil's advocate. Whether the model actually does that well is still a model-quality question, but at least the instruction is in scope.

I also appreciate that the four common research patterns (investor diligence, competitive analysis, market sizing, tech/vendor research) each have their own collection prompt. That's a small thing, but it means the agent doesn't have to guess what "good" looks like for each shape of question.

Key capabilities worth highlighting

1. Decision-shaped output structure. The 6-part default structure (exec summary → key findings → implications → risks → recommendations → sources) is the entire reason this skill exists. Most research prompts end with a "summary." This one ends with "recommendations" and "risks," which are the only two sections a decision-maker actually reads.

2. Source attribution with staleness flags. Every numeric claim is supposed to carry a source, and stale data is supposed to be explicitly labeled. In a world where LLMs confidently cite reports that don't exist, even an attempt at hygiene is a step up. The skill also has you mark estimates as estimates — which sounds trivial until you've watched an agent pass off a back-of-envelope calculation as a third-party projection.

3. The investor/fund diligence pattern. It asks for: fund size, stage, typical check size, portfolio companies, public investment thesis, recent moves, fit reasons, mismatches, and red flags. That's a real diligence checklist. If you're an angel or an early-stage founder about to take a meeting, this is the part of the skill I'd trust most — because it's mostly gathering public signals, which is exactly what an LLM can do competently.

4. The bottom-up vs. top-down TAM pattern. For market sizing, it asks for both a top-down estimate and a bottom-up sanity check based on realistic customer acquisition assumptions. It also requires you to state every logical jump as an explicit assumption. This is how real market sizing gets done; it's nice to see it encoded.

5. The quality gate before delivery. Before the agent outputs, it's supposed to verify: every number sourced or marked, stale data labeled, recommendations trace to evidence, risks included, and — the line I like best — "输出使决策更容易" (the output makes the decision easier). That's the actual success criterion.

Who should install this

Install it if you're doing any of: pre-pitch due diligence on a market, competitor mapping before a product launch, fund or angel-investor research before a meeting, vendor evaluation for a non-trivial engineering decision, or TAM estimation for a deck or internal memo. It's particularly good for solo founders, indie hackers, and analysts who need a forcing function on their own thinking.

Skip it if you need real-time data, proprietary data, or anything behind a paywall. The skill is an instruction set, not a data pipeline. If you ask it about a private company's revenue, it'll either hallucinate or refuse — the skill doesn't fix that. Also skip it if your research is purely internal (e.g., "analyze our own Q3 sales data") — it's tuned for external market intelligence, not internal analytics. And if you're already running a structured research workflow with your own prompts and a citation tool, this might feel redundant.

How to install

It's a single-file skill. Drop the SKILL.md into either ~/.claude/skills/market-research/ for a global install, or .claude/skills/market-research/ inside a specific repo. For Codex, the path is analogous under .codex/skills/. That's it — no dependencies, no scripts, no config. The repo lives at github.com/affaan-m/ECC under docs/zh-CN/skills/market-research/, and the file is plain markdown.

Concerns and limitations

I'll be honest about what I don't love.

First, the source-attribution problem isn't solved, just labeled. The skill tells the model to cite sources, but it can't verify them. If the underlying model hallucinates a Gartner report, the skill doesn't catch that — it just makes sure the citation looks sourced. Treat the citations as leads, not facts, and verify anything that matters.

Second, the quality gates are self-reported. The "check before delivery" step is the agent checking its own work. There's no external validator. On a capable model it'll do this honestly; on a lazy one it'll rubber-stamp itself. You're still depending on model quality underneath.

Third, the skill is in Chinese (zh-CN) by default. The SKILL.md I reviewed is the Chinese version. There's likely an English counterpart in the same repo, but the file path is docs/zh-CN/skills/market-research/. If you don't read Chinese, make sure you're pulling the right file. The README shows English, Portuguese, Japanese, Korean, and several other translations exist, but the skills folder path I was given is the Chinese one.

Fourth, "peaking" trend status is a signal, not a guarantee. 244K stars and 1,475 in the last week means lots of people are looking at it. That tells you it's marketed well, it solves a real pain, and the ECC ecosystem is hot. It doesn't tell you the skill itself is better than a well-written system prompt you'd write yourself in 20 minutes. The honest read: it is better than what most people will write from scratch, because it encodes a discipline (counter-evidence, staleness flagging, decision-shaped output) that people forget to include.

Fifth, no tooling. This skill is pure prompt engineering. There's no web fetch, no API call, no database lookup. If you want it to actually pull live data, you'd need to pair it with a research-tool skill or a MCP server. As shipped, it relies on the model's training-data knowledge plus whatever the agent runtime can do on its own.

Verdict

Install it. Specifically, install it if you've been burned by overconfident LLM research outputs and you want a forcing function for citation hygiene, counter-evidence, and decision-shaped output. The 244K stars are partly hype, partly ecosystem gravity, and partly a real signal that the framing works.

It's not magic, and it doesn't solve hallucination. But it changes the shape of what the agent produces, and shape is half the battle. The other half is your willingness to verify the sources it cites. If you go in with that expectation, this is one of the more useful single-file skills in the marketplace.

If you're already disciplined about how you prompt for research, you can replicate most of this yourself. But for most developers, the value is that someone already thought hard about the research process, wrote it down, and packaged it as a skill you can drop in. That's a fair trade for a few minutes of install time.

Links

• SkillsMP listing: https://skillsmp.com/creators/affaan-m/ecc/docs-zh-cn-skills-market-research
• GitHub repo: https://github.com/affaan-m/ECC/tree/main/docs/zh-CN/skills/market-research