如何使用 Claude 進行市場研究：Serena Skill 指南

It can build the model accurately; it cannot supply the numbers accurately. Ask for TAM, SAM and SOM built both top-down and bottom-up, with every filter and percentage shown, and you get arithmetic you can check and argue with. The published figures it anchors on are recalled from training data and may be stale or wrong, so the discipline is to make it mark every number as known, published or assumed, then verify the published ones against the statistics office, industry body or company filing it names.

Because an unguided model fills gaps with plausible fiction and states all of it at the same confidence. It has no way to signal that one figure came from a real report and the next was chosen to make the paragraph work. The fix is not asking it to be accurate; it is forcing it to tag its own sources and flag its own estimates, so the guesswork arrives labelled as guesswork.

Five things: the product described in one concrete sentence, the geography and the year, every real figure you already have such as your pricing or conversion rate, the competitors you already know, and what the research is for. Without the geography and time frame you get a flattering global number; without your real figures the model has to invent more of the model than it should.

Not for a private company, and neither can anyone else from outside. Any specific revenue figure for a private competitor is an estimate presented as a fact. Ask for competitor claims about pricing, funding or revenue to be marked VERIFY with the page you should check them on, and treat anything unsourced as a hypothesis rather than a number to put in a deck.

It cannot read the paywalled Gartner, IDC or Forrester reports that anchor most market sizing, it has a training cutoff so recent funding rounds and new entrants may not exist for it, it cannot speak to your customers, and statistics such as confidence intervals and representative sampling are arithmetic, which is where language models quietly fail. It is a research scaffold, not a research department.

It is a ZIP file with a SKILL.md at the root. In the Claude desktop app, open Customize then Skills, add a skill, upload the ZIP, and toggle it on. Skills work on Free, Pro, Max, Team and Enterprise; the real requirement is code execution, which you enable as 'Code execution and file creation' under Settings then Capabilities. On Team and Enterprise an owner can disable it for the whole organisation. In ChatGPT or Gemini, open the SKILL.md and paste its contents into custom instructions instead. Every KissMySkills purchase carries a 30-day money-back guarantee.