Writing Agent Skills for Digital Marketing: 6 Rules From 30 Production Skills

I'll never forget the morning I realized my best-performing ChatGPT conversation wasn't going to scale. I had spent 40 minutes tuning a single prompt that turned a one-paragraph brief into a 12-email lifecycle sequence — and by the next week, I'd pasted it into a new chat three times, and all three runs came out slightly different.

That was a prompt masquerading as a workflow. The workflow — what I actually wanted repeated — belonged in a skill.

If you've been working with AI agents in 2026, you've heard the word "skill" thrown around a lot. Anthropic launched Claude Skills in October 2025 and open-sourced the format in 2026; Andrew Ng put out a course on it; every agent builder now has a "skills system." For marketers, this is the part of the agent stack that actually pays off: a skill is the unit of reuse — it's how you package a workflow you trust so the agent can run it again tomorrow, on a different client, without you retyping the same system prompt.

I've now authored or co-authored more than 30 skills for digital marketing — including the ones this blog runs on (writing posts, generating cover images, planning a 90-day content calendar). Most of what makes them good is unglamorous: discipline about scope, clarity about when the skill should fire, brutal honesty about what the agent will get wrong. Here's the short version.

1. Write the frontmatter last, not first

A skill has two parts: a YAML frontmatter (the interface) and a Markdown body (the work). Most people, including me on day one, want to start with the body — the actual instructions. Don't. Write the name and description last.

The description is the trigger. It's the one piece of metadata the agent loads without reading the body. When an inbound asks "draft me a welcome series," the agent scans skill descriptions and picks. A vague description ("writes marketing emails") loses to focused ones ("drafts a 7-email SaaS welcome series, 14-day cadence, with the HML framework explicitly named in each subject line").

Here's the test I now use: would a competent junior marketer reading just this 1-sentence description know when to use this skill and when not to? If they say "only when..." write the boundary in.

2. One workflow per skill

The single biggest mistake in early skill design is bundling. People want "an SEO skill" that does keyword research, on-page optimization, internal linking, and content refresh. They write a 4,000-word SKILL.md and wonder why the agent loses the plot by step 11.

Split it. A skill should be one workflow (input → transformation → output), even if the underlying capability is broad. If your "SEO" workflow has a branch where you do keyword research vs. one where you do internal linking, those are two skills. Each one becomes 200–600 lines of focused Markdown, instead of a single bloated file. The agent reasons about them better. You debug them better. You can hand one to a colleague without dragging the rest along.

For marketing work, a useful working definition: if a workflow has a different output artifact (a content brief vs. an internal link plan vs. a content refresh ticket), it's a different skill.

3. Bake in the brand. Specifically, your client's voice

Generic skills age out within a season. The 12-email lifecycle workflow I wrote in February was useful, mostly. The same workflow, with 200 words about a brand's voice — Forbidden words: "synergy," "ecosystem," "unlock." Preferred openers: "Quick question —", "I noticed..." — produced emails that sounded like the brand instead of like an LLM.

The pattern is to keep the workflow instructions universal in SKILL.md, but stash brand-specific inputs (do/don't lists, exemplar paragraphs, product taxonomy) in a sidecar file the skill loads conditionally. That way one skill can serve five clients without forking the code.

This is also how you skip the "AI smell" — the cadence patterns and cliche words that mark AI-written copy. Codify your own.

4. Use scripts for the deterministic parts, prompts for the judgment parts

This one took me longer than I want to admit. Inside a marketing skill there are pieces that should never be re-decided (the JSON-LD schema template, the RSA headline pin counts, the canonical UTM parameter order) and pieces that genuinely benefit from the model's judgment (the email subject line, the meta description, the hook). Don't ask the LLM to recompute the first set every run. Hand it a script — Python, a shell snippet, even a Notion formula — and let the prompts handle the parts where the cost of variability is zero.

This shrinks tokens, sharpens outputs, and makes failure modes far easier to debug. When a skill breaks, a deterministic script breaks loud (a stack trace). A prompt fails quietly (a slightly worse email subject line).

5. Document the failure modes

This is the rule nobody writes about. A good skill tells the agent what to do when things don't match the happy path. No GSC access? Then the SKILL.md says fall back to this format. Numbers below 50? Then output this disclaimer. Tool call rate-limited? Then queue and retry, don't loop.

The failure-mode section is also where you capture team knowledge. "If the brand uses a UK-English style, swap US/UK spellings here." "If the input is a YouTube URL, always fetch the transcript; if it's a podcast RSS feed, fetch the latest 5 episodes; if it's an article URL, fetch and follow the first 3 outbound links." These are the kind of decisions humans make automatically that an LLM won't unless you pin them.

6. Version the skill. Treat it like code

Skills are living artifacts. A skill I wrote 4 months ago has been edited 20+ times — after I saw a real failure, after a new model changed its defaults, after a tool API shifted. The folder structure (SKILL.md + scripts/ + references/) makes this easy: diff the Markdown, save versions of the scripts, retire outdated references.

I tag skills with a last-updated date in the frontmatter and keep a one-line CHANGELOG note inside the SKILL.md. The cost is a few seconds per edit; the upside is that when a colleague inherits the workflow six months later, they can read what's been tried, dropped, and replaced.

The minimum viable marketing skill

If you only build one skill this quarter, build the one that does the job you keep doing manually. For me that's content production. For a paid-media team it might be PMax asset refresh. For SEO, it's a 4-step content brief pipeline.

A skill doesn't need scripts. It doesn't need references. It just needs a sharp frontmatter description, a clear workflow in the body, and a name that's honest about what it does and isn't. Start there.

Once you have one, you'll write three more, because that's how AI work actually scales. Not by adding more tabs to ChatGPT. By packaging the workflows you've already won.