Ask a few people on your team to draft the same follow-up email and you'll get a few different results. Ask your best person to do it twice, a week apart, and you might get two different results from them too. That's the real problem with relying on prompts. Quality depends on who's typing and how well they remember what worked last time.
Claude skills fix that. A skill is a pre-built instruction set Claude reuses every time, so the result doesn't depend on memory or how good the prompt happens to be that day.
What a skill actually is
Think of a skill as a standard operating procedure, but for Claude instead of a person. Most businesses already have an SOP for onboarding, for handling a complaint or for running a client meeting. A skill is the same idea. Instead of telling a person how to do something, you're telling Claude.
Without a skill, output quality depends entirely on how well the person prompting Claude knows what they're doing. Your best prompter gets great results. Someone new to the tool gets something generic. Even your best prompter gets inconsistent results if they're typing from memory each time.
A skill takes that expertise out of one person's head and locks it into the tool. The knowledge travels with the workflow, not with the individual.
Choosing the right task for your first skill
Pick something your team does repeatedly, not a one-off. If someone on your team gets asked the same type of question more than three times, that's a skill waiting to be written.
Good candidates: drafting a specific type of email, generating a sales proposal, reviewing work against a set of criteria or generating a report from raw input.
The four components every skill needs
Once you've picked the task, every skill is built from four parts.
The trigger. What tells Claude to use the skill. This is a description of the situations or phrases that should activate it, for example: use this skill whenever someone asks to draft a follow-up email after a client call. Be specific. A vague trigger means Claude won't know when to apply it.
The instructions. The step-by-step breakdown of what Claude should actually do. This is the heart of the skill. Think about what your best team member does naturally when they approach this task and write it down. What do they check first? What do they always include? What's the order?
The output format. What the result should look like: a draft email or a bullet point summary, not a task assignment. Don't leave this open-ended. If you don't specify, Claude will make its own call, and that's where inconsistency creeps back in.
Guardrails. What Claude should never do in this context. Especially important for anything client-facing: don't make up specific numbers, don't include pricing without checking, always flag if something is missing rather than guessing.
Building a skill: a live example
For the demo, Warwick built a skill to process a sales call transcript and update Pipedrive. He prompted Claude directly: take a transcript after a sales call, process it, update the deal in Pipedrive and identify next actions.
The prompt told Claude where the transcript would be located, how it's named and exactly what to do with it: find the deal in Pipedrive (being specific about what kind of deal to look for), update the activity, and only add to the existing description rather than overwrite it. It also told Claude the next steps needed to keep the sales process moving, and to suggest updates to any relevant empty fields on the deal.
Being specific here matters. It leaves less room for Claude to interpret the task its own way and tightens the scope to exactly what needs to happen.
Reviewing and saving Claude's draft skill
Once Claude finishes building the skill, it comes back with the workflow it's proposing. Read through it and check it's accurate. If there's ambiguity, or if it hasn't gotten something quite right, keep chatting with Claude in the same conversation and ask it to change the approach.
Claude opens the skill for you to read in full at this point. That's your chance to catch anything hallucinated or inserted that you don't actually need, and ask Claude to remove it. Once you're happy, save the skill and it's added to your skill library, ready to use.
Why you need to refine a skill after you build it
This is where most teams drop the ball. They build the skill, it's ready to go, and they never come back to ask whether it's working the best way it could.
Run some real inputs through the skill, look at what comes out and give Claude feedback. Claude can update the skill for you directly. A new skill is always version one. Just like app updates, every revision makes it better.
Two ways to run a skill
Once a skill exists, there are two ways to use it. The first is natural language: tell Claude what you want in plain terms, for example, “I've just had a call with Paul and I'd like to process the transcript so I can update the deal in Pipedrive.” Claude figures out which skill to use from that.
The second is invoking it manually with a forward slash, then choosing it from the list or typing its name. You can hit return and let Claude prompt you for details, or provide the details upfront, like naming the specific call you want processed. Skills also show up in the dropdown menu.
Real example: the proposal generator
This is the one with the clearest return. Writing a proposal used to take half an hour or more. The skill gets it done in five minutes or less.
It pulls details from the Pipedrive deal and the sales call transcript, then combines them into a proposal structured the way MinorCo structures its proposals, using a template document. It asks for clarification on things like pricing tiers or what should be included, and can suggest additions the user might not have thought of.
It gets you about 80% of the way there. There's still a review pass: personal touches, a check that nothing reads oddly. But the time saved getting to that 80% mark is worth it on its own.
Real example: the daily triage
The daily triage pulls from Gmail, Pipedrive and Asana and gives a consolidated view of what's on your plate for the day.
Getting the output right takes some upfront work, particularly deciding exactly what you want it to look like. If you'd rather not build that yourself, MinorCo offers to set it up as part of its consulting service.
Your challenge
Find a process that's taking time every week, something different team members handle inconsistently, and turn it into a skill. Don't let the fear of an imperfect first version stop you. Take the time to figure out what works. You'll thank yourself later.
Watch the full walkthrough to see the skill built end to end, including the exact prompt Warwick used.
FAQ
What is a Claude skill? A pre-built instruction set that tells Claude how to handle a specific repeated task, the same way an SOP tells a person how to handle it. It's made up of a trigger, instructions, an output format and guardrails.
How is a skill different from writing a prompt each time? A prompt depends on whoever's typing it and how well they remember what worked last time. A skill locks that approach into the tool, so results stay consistent no matter who runs it.
What should be in a skill's instructions? The step-by-step approach your best team member takes naturally: what they check first, what they always include and the order they do things in.
How do you run a skill once it's built? Either in natural language, by describing what you want and letting Claude match it to the right skill, or manually with a forward slash followed by the skill's name.
Can a skill update a CRM like Pipedrive? Yes. One of the examples in the video is a skill that reads a sales call transcript and updates a Pipedrive deal, including the activity and next steps, without overwriting existing data.