Claude skills vs custom instructions — which should you use?

People put the wrong things in the wrong place constantly. Then they wonder why their agent is inconsistent.

Here’s where the line actually is.

Custom instructions are your baseline

They load in every conversation. Every one. Use them for things that should be true regardless of the task:

• Preferred response length and tone
• Coding language and style conventions
• Formatting preferences
• Things you never want Claude to do
• Background context about who you are

Think of them as standing orders. They don’t change between tasks — they set the frame for everything that happens inside a session.

Skills are for specific jobs

A skill is invoked for a particular type of work. It defines:

1. What task it handles
2. What inputs it needs before starting
3. The step-by-step process to follow
4. What the output should look like

That structure is why skills outperform prompts for complex repeated work. A prompt tells the model what you want. A skill tells it how to get there.

The test is simple

If you catch yourself typing the same instruction in every session, it belongs in custom instructions. “Always explain your reasoning before answering.” “Use TypeScript.” “Keep responses concise.” Those are preferences — they apply everywhere.

If the instruction only makes sense for a specific type of work, it belongs in a skill. “When reviewing code, check X then Y then Z” — that’s a process, and processes belong in skills.

One question that cuts through the confusion: is this a preference or a process?

• “Always use British English” → custom instructions
• “When reviewing code, check X then Y then Z” → skill

The mistakes that cost people the most time

Putting everything in custom instructions. A 2,000-word custom instruction block that tries to cover every workflow degrades response quality. The model has to hold all of it in context all the time — even when 90% of it isn’t relevant to the current task. Shorter, focused custom instructions consistently outperform comprehensive ones.

Skills that are too broad. “Be a great marketer” is not a process. A skill needs to say what inputs it requires, what steps it follows, and what the output looks like. Narrow skills that handle one workflow extremely well beat vague skills every time.

Duplicating the same instruction in both places. If “British English” is in custom instructions and also embedded in your writing skill, you’re adding noise. Pick one home.

Using a skill for one-off tasks. Skills are for repeated work. If you’re doing something once, a well-written prompt is faster. Invest in a skill when you know you’ll run the same workflow again and again.

They work better together than apart

Custom instructions set the style. Skills add process on top of it.

A developer might have custom instructions that set their language preferences and communication style — then skills for PR review, systematic debugging, and API critique. Each skill inherits the tone from custom instructions but adds its own structure on top.

That’s the right layering. The style is global. The process is per-task.

Practical setups by role

Developers: Custom instructions covering language preferences, code style, communication style. Skills for code review, debugging workflows, API design. The custom instructions filter the style; the skills determine what gets checked and in what order.

Marketers: Custom instructions covering brand voice, target audience, channels. Skills for keyword research, content briefs, campaign analysis. Same pattern — style is global, process is per-task.

Founders: Custom instructions covering communication style, company stage, strategic context. Skills for board prep, user feedback synthesis, competitive teardowns.

Browse skills by audience to see what’s available for your role.