What is a Claude skill? How we're turning agency processes into AI skills
Last updated: 17 November 2026
Most businesses using AI still treat it like a clever intern: you explain the task from scratch, it does a reasonable job, but then next week you end up explaining it all again. A Claude skill fixes all that. It packages your process, templates, and quality checks into something the AI loads automatically whenever the task comes up. This means the work is done your way every time.
As a member of Anthropic's Claude Partner Network, Refuel has started turning its own agency deliverables into skills, and the results are worth sharing.
What is a Claude skill?
A skill is a folder containing instructions and resources that Claude can use when they're relevant. At minimum it has a single file, SKILL.md, with two pieces of metadata at the top:
- Name: a short, lowercase identifier. Oddly enough, Anthropic doesn't allow the words "claude" or "anthropic" in it.
- Description: what the skill does and when to use it. This matters more than anything else, because Claude reads the description to decide whether to load the skill for a given request.
Below that sit the instructions themselves: the steps, the standards, examples of good output and anything the AI tends to get wrong. A skill can also include reference files, templates and scripts.
Why skills don't bloat the AI's memory
Anthropic designed skills to load in stages. Only the name and description of each skill are loaded up front, at around 100 tokens each. The full instructions load only when the skill is triggered, and reference files or scripts are read only when needed. Scripts run in the background and only their output enters the conversation. That means you can install dozens of skills, each with detailed checklists and templates, without slowing down everyday conversations.
Where Claude skills work
Skills work in the Claude apps (including Cowork), in Claude Code and through the Claude API. Anthropic also provides ready-made skills for creating Word, Excel, PowerPoint and PDF files. One catch: custom skills don't sync between surfaces. A skill uploaded to the Claude app needs to be added separately to Claude Code or the API.
How we are using Claude skills
We set ourselves a simple goal: build a skill for as many repeatable deliverables as possible, based on the processes and quality standards we already use. The first set covers:
- SEO audits: technical, on-page, backlink, keyword and AI search checks, pulling data from Semrush and Search Console.
- Weekly paid media reviews: Google Ads and Meta Ads checks, pacing, recommended changes with reasons and a change log, with nothing applied until a person approves it.
- HubSpot implementations: portal audits, discovery, solution design, migration mapping, testing scripts and training guides.
- Website QA: go-live checklists, accessibility, tracking, SEO and speed checks with a defect log ranked by severity.
- Email retention audits: Klaviyo setup, deliverability, flows, forms and campaigns, scored with next steps.
- Project prep and quoting: WIP meeting agendas from project data, and quotes built from our product library and quoting rules.
The point isn't to remove people from the work. It's to make sure every audit, review and quote starts from the same standard, regardless of who runs it, and gets there faster.
Why we joined the Claude Partner Network
Refuel was accepted into the Claude Partner Network in June 2026. Anthropic launched the network in March 2026 for organisations that help businesses adopt Claude, backed by an initial US$100 million commitment to partner training, enablement and market development. Partners get access to Anthropic's training and certification paths, enablement materials and technical support.
For us it reflects a simple shift. We've spent years as a marketing, CRM and web agency, and AI now runs through all of that work. Our team has worked through the partner learning path, and we apply what we learn to our own delivery first: skills, connectors and automations that we use every day before we recommend them to anyone else. That means when we help a business bring AI into its marketing, sales or operations, the advice comes from building and running it ourselves, not from a vendor demo.
What we have learned building them
1. Skills are only as good as the process behind them
Every skill we've built started from a documented process. If your process lives in someone's head, writing the skill forces you to write it down, which is useful in itself. If you don't have a process, a skill will just automate inconsistency.
2. Pair skills with connectors
A skill tells the AI how to do the job. A connector (an MCP server) gives it access to the data. Our SEO audit skill is far more useful with the Semrush connector than without it, and our HubSpot work leans on HubSpot's own MCP server. When no connector exists, we've written fallbacks into the skill, such as using the browser to read Analytics or Search Console.
3. Build in human checkpoints
For data-heavy work like an SEO audit, a fully automated run hides too much. The AI filters and sorts large data sets, and unless a specialist can see what it discarded, they can't properly check its conclusions. Our better skills run in rounds: the AI does a pass, a person reviews and adds direction, then the AI continues. Saving a version after each round also shows exactly how much the human input changed the result.
4. Write the description carefully
If a skill doesn't trigger when it should, the description is almost always the problem. Spell out the phrases people actually use ("run an SEO audit", "check this site before go live") and be explicit about what the skill doesn't cover.
5. Expect to merge and refine
Once a team starts building skills, you'll quickly find two people have built versions of the same thing. That's a good problem. Compare them, keep the best parts of each and settle on one version so the whole team works the same way.
How to use Claude skills in your own business
You don't need to be an AI marketing agency or a developer to get value from skills. Good candidates are tasks that are repeated often, follow a known standard and currently depend on one or two people to get right. For marketing teams, the most useful Claude marketing skills tend to cover:
- Writing campaign briefs or monthly reports in your house format
- Checking landing pages before launch
- Applying brand voice and style rules to copy
- Preparing board or leadership updates from the same data sources each month
Claude can help write the first draft of a skill. Anthropic offers a skill creator that interviews you about the task and produces the files. From there, test it on real work, note where the output falls short and add those lessons to the instructions.
A note on security
Anthropic's own guidance is to treat skills like installed software. A skill can instruct the AI to run code or call tools, so only use skills you've built yourself or obtained from a trusted source, and review every file in a skill before installing it. Be especially careful with skills that fetch content from external websites.
Turning know-how into repeatable systems
The businesses getting real value from AI aren't the ones with the cleverest prompts. They are the ones turning their know-how into repeatable systems: clear processes, the right data connections and checks in the right places. Skills are one of the most practical ways to do that, and they reward the work you've already put into documenting how things should be done.
Want help putting AI to work in your marketing?
Refuel Creative helps Australian businesses with HubSpot, SEO, paid media, websites, RevOps and AI. As a Claude Partner Network member, we can help you work out where Claude, skills and connectors fit in your marketing, sales and operations, then build them properly. To see what that could look like for your business, talk to our team.