如何將 Claude 用作財務分析師：Miles Skill 指南

Yes, with one important limit. Claude is strong at the structure and the reasoning: laying out a DCF, choosing and justifying a peer set, spotting quality-of-earnings flags in a set of statements, and explaining what is driving a result in language a non-finance reader can act on. It is weak at being a calculator. Language models predict text, so long chains of compounding and multiplication are where errors creep in, and the output still looks like a spreadsheet when it is wrong. Use it to build and explain the model, then re-run the arithmetic yourself before anyone relies on it.

A skill is a single file you load into Claude, ChatGPT or any AI chat that turns it into one specialist for the whole conversation. The Miles financial analyst skill makes the model list its assumptions separately from your figures before it answers, build a base case and a downside case rather than one number, and mark every estimate so you know what to verify. Those rules apply to every follow-up question, including the casual ones where you would have forgotten to paste a prompt.

$29, one time. No subscription and no seat fee: you download the file and keep it. If you want the wider finance stack, the Finance Agent Bundle is $99 for all five finance agents against $160 bought separately.

It can build the structure and the logic, and that is most of the work. It will lay out the growth path, the margin forecast, the WACC inputs, the terminal value and a sensitivity table. What it cannot do is know your real numbers or today's market data, so it fills those gaps with estimates. On a typical valuation, four of the six key inputs are the model's choice rather than yours. The point of using a skill is that those four are listed openly instead of buried in the answer, so you can replace them.

That depends on your provider settings and your own policy, not on the skill file. Check whether your plan trains on your conversations and whether your company allows the data to leave its systems. As a practical habit, many people run the structure on anonymised or rounded figures first and only paste the real numbers once they trust the shape of the output.

Yes. The file is plain Markdown, so the method travels. In Claude you upload it through Customize then Skills, or paste it into project instructions. In ChatGPT, Gemini, Grok or Copilot you paste the contents into custom instructions or a system prompt and it behaves the same way.

Four things. Re-run the arithmetic in a spreadsheet. Replace every figure marked as an estimate with your own data where you have it. Source any trading multiples yourself, since comps move weekly and anything recalled from training is stale. And sanity-check the peer set, because a comparable that is twice the size or in a different regulatory regime will quietly drag the whole valuation. The output is structure and explanation, not financial advice.

A good prompt gets you most of the behaviour once. The skill gets it every time, without you re-typing the rules, and it carries the parts a prompt does not: the earnings-quality red flags, the peer-selection logic and the investment memo format. In finance the risky answer is usually the casual follow-up asked three messages in, which is exactly where a pasted prompt has stopped applying.