When to use Claude projects vs skills, and how to avoid overloading the AI

Executive overview

Every active skill loads its title and description into every conversation — stack too many and Claude picks the wrong one. Projects scope the AI to a single activity; skills encode a reusable process or format.

The default should be projects; reach for skills only when the same procedure must run across many different projects.

Two questions that tell you if you need projects or skills

• Do I need this kind of work more than once?
• Does quality have to be high every single time?
• If yes to both, projects or skills are worth building — otherwise, plain chat is enough.

What makes a good project

• One project = one specific activity, not one project per client or company.
• Wrong: dump all ACME files into a single project. Right: separate projects for client updates, monthly closes, proposals, contract review.
• Focused context means the AI is not distracted by unrelated files.
• Projects scale freely — hundreds can coexist because each opens in isolation.

Projects in the browser vs Claude Cowork (desktop)

• Browser projects: knowledge files uploaded explicitly; system instructions set in the project UI.
• Cowork projects: a folder on disk; CLAUDE.md inside the folder is the system instruction; all subfolders are the context.
• Open a subfolder in Cowork and the AI sees only that subfolder — not the parent or siblings.

What makes a good skill

• Use a skill when the same procedure, analysis steps, or output format must be consistent across many contexts.
• Examples: proposal writing with fixed branding, a financial close checklist applied to multiple clients.
• Skills contain a title, one-line description (always loaded), detailed instructions, and optional reference files (loaded selectively).

Skills in the browser vs Cowork

• Browser: all enabled skills are visible to the AI in every conversation — the AI proactively calls whichever seems relevant.
• Cowork desktop: skills can be pinned to specific subfolders, so only the relevant skills are active inside that folder.
• This means skills scale on the desktop but not in the browser — too many browser skills overwhelm the AI.

Two patterns for combining projects and skills

• Pattern 1 — skill alone: open a new chat anywhere, call the proposal-writer skill, drop in the prospect context, get the proposal, close the chat.
• Pattern 2 — skill inside a project: a client reconciliation project holds client-specific rules; the project instructions also call the financial-close skill for the standard process. The two layers merge on each run.

How to create a skill

• Do not write a skill from scratch.
• Have a full conversation with Claude until you get an output you are happy with.
• At the end of that conversation, paste the prompt: "Use the skill creator to create a skill based on the above — strip out client-specific inputs, keep the process, standards, and format."
• Save the resulting skill file.

Three reasons skills are worth building

• Portability: a skill downloaded from Cowork can be uploaded to ChatGPT and works equally well — no model lock-in.
• Chaining: project instructions can call multiple skills in sequence (step 1: skill A, step 2: skill B).
• Reuse: one skill (e.g. proposal writer) serves every project without being locked to any single one.

Best practices and limits

• Beginners should start with projects, not skills — simpler use cases don't need a skill to get good outputs.
• Browser skill limit: keep active skills to 13–15 maximum to avoid confusing the AI with similar titles.
• Explicitly call a skill with a / slash command rather than relying on the AI to pick the right one automatically.
• Slash commands are not foolproof — if too many skills exist, the AI may still proactively call the wrong one.
• Definitive fix for skill overload: use Cowork and pin each skill to the relevant subfolder.