如何使用 Claude 撰寫部落格內容：Cleo Skill 指南

Not for being AI-written. Google's spam policies target scaled content abuse, defined as when many pages are generated for the primary purpose of manipulating search rankings and not helping users, and the policy says this applies to unoriginal, low-value content no matter how it is created. Using a model to write is not the offence; producing volume with nothing in it is, and a human typing it does not save it. The useful question about a draft is whether anyone would be worse off if the page did not exist.

Because of a small set of recognisable habits rather than anything mysterious. The throat-clearing opening about today's fast-paced landscape. An intro that restates the title the reader already clicked. Lists that are always three items long. The not-just-X-it-is-Y construction. Stacked hedges such as can often be one of the more effective approaches. Uniform paragraph length and rhythm. A conclusion that recaps. And the surest tell of all, paragraphs containing no proper noun, number or date.

It can produce the shape of expertise without any of the content of it. It has read thousands of posts about your subject and none of your work, so it does not know what your customer said on the call, what the number actually was, which supplier let you down, or what you tried first that failed. Those are the parts readers remember and the parts that earn links. The five-second test per paragraph: could this sentence have been written by someone who has never done the thing?

Write a brief rather than naming a topic. It should contain who is reading and what they already believe, what should change in their head, the position you are claiming that a reasonable person could disagree with, what you know that research alone could not supply, what would make the post wrong, and the search phrase it should earn. That fifth field, the conditions under which your position does not hold, is the one almost nobody fills in and the one that stops the draft hedging.

No. A model asked for a whole post optimises for finishing it; asked for one section with the evidence in hand it optimises for that section, and the quality difference is larger than any prompt wording trick. Approve an outline first, drafted as an argument with a claim per section rather than a list of headings, then draft section by section with the source material pasted in and an instruction to mark a missing fact rather than invent one.

Run a deletion-only pass, in its own conversation, with no permission to rewrite or improve anything. Cut the first paragraph unless it makes a claim, every sentence that says what the post will do or just did, every hedge that does not change the meaning, every adjective that survives its own removal, and the concluding recap. Target at least 25 percent shorter. A model asked to both improve and cut will improve, because improving is easier and looks like more work.

Give it the material and forbid it from supplying anything else. The instruction that works is to have it write a placeholder such as NEED: what whenever a sentence requires a fact you did not provide, rather than filling the gap. Asked for concrete detail without source material, a model produces plausible statistics, studies and quotes that do not exist, which is the most dangerous failure on a published page.

The download is a ZIP with SKILL.md at the root of the archive, not inside a nested folder. In the Claude desktop app open Customize then Skills, upload the ZIP and toggle it on. Skills need code execution enabled, under Settings then Capabilities, or by an owner in organisation settings on Team and Enterprise. Anthropic's help centre currently lists Skills on Free, Pro, Max, Team and Enterprise, while its Academy tutorial lists Pro, Max, Team and Enterprise, so if you are on the free plan check Settings then Capabilities for your own account rather than trusting either page. In ChatGPT or Gemini, paste the contents of SKILL.md into custom instructions instead.