How to Write a Claude Skill for Marketing
Writing your first Claude Skill is easier than it sounds, because a Skill is mostly just clear documentation of how you want a task done. If you can brief a new hire, you can write a SKILL.md. The difference between a Skill that works and one that disappoints is not technical polish; it is the clarity of the description that triggers it and the specificity of the instructions inside. Here is a step-by-step way to write one for a real marketing task, using a brand-voice Skill as the worked example.
Step 1: Pick one repeatable task
Do not start with something sprawling. Pick a single task you already do often and re-explain to AI every time, brand voice, a specific content format, ad copy for your offers, a reporting layout. The best first Skill is the one that will save you the most repeated context. For most marketers that is brand voice, because it touches everything, so we will build that as the example. A good Skill has a clear, bounded job; if you cannot describe the task in a sentence, it is too big to start with.
Step 2: Write the description (this is the make-or-break part)
Every SKILL.md opens with a short description, and it does more work than any other line in the file, because Claude uses it to decide when to apply the Skill. A weak description ("brand stuff") means the Skill never triggers at the right moment or triggers at the wrong one. A strong description states plainly what the Skill does and when to use it: for example, "Applies the Opère18 brand voice, tone, and banned-word rules to any marketing copy, use whenever writing or editing posts, emails, ads, or landing-page text."
Write the description for the trigger, not for a human reader. Name the task and the situations it covers in the words a request would actually use. If your Skill should fire whenever you write social posts, make sure "social posts" appears. The clearer the description, the more reliably the Skill activates on its own, which is the entire point.
Spend more time on the description than feels reasonable. It is the difference between a Skill that activates itself at the right moment and one that sits unused because Claude never knows it applies.
Step 3: Write the instructions like a brief for a sharp new hire
The body of the SKILL.md is where you document how the task is done. Write it the way you would brief a capable new hire who knows marketing but not your brand: specific, concrete, and ordered. For a brand-voice Skill that means the actual rules, the tone in a few adjectives, the sentence patterns you favour, the words and phrases you never use, the perspective and beliefs behind the voice, and a short do-and-do-not comparison. Vague guidance ("be professional but friendly") produces vague output; specific rules ("short sentences, no em dashes, never say synergy, lead with the problem") produce your voice.
The strongest instructions include examples. A couple of before-and-after rewrites, or two or three lines that unmistakably sound like you, teach the standard faster than paragraphs of description. Show the Skill what good looks like, not just tell it.
Step 4: Add resources only if the task needs them
A brand-voice Skill may need nothing but the SKILL.md. Other Skills benefit from bundled resources: a content Skill might include your post template and SEO checklist, a reporting Skill a spreadsheet layout, an audit Skill the exact criteria to check against. Add these only when the task genuinely requires them. The discipline is the same as good documentation, include what is needed to do the job right, and nothing that just adds noise.
Step 5: Test it on real work and refine
A first draft of a Skill is a hypothesis. Put it to work on genuine tasks for a few days and watch where the output drifts from what you wanted. Every miss is information: if the voice comes out too formal, your tone rules were not specific enough; if it uses a banned phrase, add it to the list; if the Skill did not trigger when it should have, sharpen the description. Fold each correction back into the file. Within a week of small refinements, a Skill goes from roughly right to reliably on-standard.
This feedback loop is the real work, and it is worth it, because a refined Skill keeps paying off on every future task with almost no prompting. You are not writing a document you file away; you are training a capability you will use daily.
Common mistakes that make a Skill fail
Most disappointing Skills fail in a handful of predictable ways, and all are avoidable. The first is a vague description, so Claude never knows when the Skill applies and it sits unused; the fix is to name the task and the situations in the words a real request would use. The second is instructions that are too general, "write engaging content" teaches nothing, where specific rules and examples teach your standard. The third is trying to make one Skill do too much: a single Skill that handles voice and SEO and formatting and research becomes muddy, where several focused Skills stay sharp and compose cleanly.
Two more are worth naming. Skipping examples is a quiet killer, because a couple of before-and-after rewrites teach the standard faster than paragraphs of description ever will. And treating the Skill as finished after the first draft guarantees mediocrity, because the real quality comes from the refinement loop, testing on real work and folding corrections back in. A Skill you write once and never touch again is a first draft masquerading as a system.
The through-line: a Skill is only as good as the clarity you put into it. The technology is forgiving; vagueness is not. Sharp description, specific instructions, real examples, one bounded job, and steady refinement, get those right and the Skill works reliably. Miss them and no amount of tooling saves it.
If you want a checklist to keep beside you as you write, it is short. Can you state the task in one sentence? Does the description name the situations a real request would use? Are the instructions specific enough that a stranger could follow them and sound like you? Did you include at least one example of good output? Have you tested it on real work and folded the corrections back in? Five yes answers and you have a Skill that will earn its keep every time you use it. That is the entire bar, and it is well within reach of any marketer, no coding required.
Then do it again
Once your first Skill is solid, you have both a working capability and a template. The next time you notice yourself re-briefing AI on the same task, write it into a Skill the same way: one bounded job, a sharp description, specific instructions, examples, test and refine. Do that a handful of times and you have a marketing Skill library, the compounding asset that turns AI from a tool you prompt into a system that runs your marketing to your standard.
FREQUENTLY ASKED
How do I write a SKILL.md for a marketing task?
Pick one repeatable task, write a clear description stating what the Skill does and when to use it, then document the instructions like a brief for a sharp new hire, specific rules, your standards, and a couple of examples. Add resources like templates only if the task needs them, then test on real work and refine until the output is reliably on-standard.
What makes a Claude Skill work well or fail?
Two things. The description, because Claude uses it to decide when to apply the Skill, so a vague one means it never triggers correctly; and the specificity of the instructions, because vague guidance produces generic output while concrete rules and examples produce your standard. Polish matters less than clarity of trigger and precision of instructions.
What should the first marketing Skill I build be?
Usually brand voice, because it touches every piece of output and is the context you most often re-explain to AI. It is bounded enough to write quickly and high-leverage enough that getting it right improves all your content, ads, and emails at once, which also makes it the ideal template for your next Skill.
How long does it take to get a Skill working well?
A basic version takes an hour or two to write. Getting it reliably on-standard takes a few days of testing on real tasks and folding each correction back into the file. The refinement loop is the real work, but once a Skill is dialled in it keeps producing on-standard output on every future task with almost no prompting.
RELATED SERVICES
Want this done for your business?
Free audit. No pitch. 24-hour turnaround.