MCPFlixFree Beta

Mental Model · 7 min read

Skills vs MCP vs Prompts: the five-tool carpentry framework

Claude has five tools in its belt. Most teams reach for whichever one they used yesterday. Here is when to use which — without overthinking it.

The framework comes from Tahir's January 2026 essay Agent Skills vs MCP vs Prompts vs Projects vs Subagents. He pitches it as carpentry: a hammer is not a screwdriver, a screwdriver is not a saw. Reach for the wrong one and the work still gets done — but you bleed time, tokens, and money you did not need to spend.

The killer heuristic

If you find yourself typing the same detailed prompt over and over, that is a signal you should make a Skill.

— Tahir, 2026

The five tools, side by side

Here is the whole framework on one screen. Read the metaphor column first — it is the part you will remember in six months.

ToolMetaphorWhen to use
PromptWhat you say right nowOne-off, ephemeral — "rewrite this paragraph"
SkillTraining manualSame task done the same way every time
ProjectA dedicated workspaceAn initiative with its own background facts (200K context + RAG)
SubagentSpecialized workerIndependent task, isolated context, restricted tools
MCPUniversal plugLive access to an external system (Slack, Postgres, Jira)

Prompts: the one-shot

A prompt is what you say in the chat box right now. It vanishes when the session ends. Use prompts for genuinely one-off work — rewording an email, asking a question, sketching an idea. The mistake is using prompts for things you do every Tuesday.

Pricing tier: free, every Claude user gets them.

Skills: the SKILL.md

A Skill is procedural knowledge written down. It lives in ~/.claude/skills/<name>/SKILL.md as Markdown with YAML frontmatter that tells Claude when to load it. The metadata (about 100 tokens) is always in context; the body (under 5,000 tokens) only loads when triggered. Files in scripts/ and references/ load only when the body asks for them.

---
name: pr-review
description: |
  Use when the user asks for a code review on a pull request,
  diff, or branch. Do NOT use for greenfield design feedback.
---

# PR Review

When invoked:
1. Run `scripts/lint-diff.sh`
2. Read `references/style-guide.md` for repo conventions
3. Output review as inline comments + a summary table.

Pricing tier: free for all users. Skills travel with you — drop the folder into another machine and it works.

Projects: the workspace

A Project is a dedicated workspace on claude.ai with its own context window (up to 200K tokens) and its own private knowledge base. Drop in PDFs, transcripts, or repo files; they are searchable across every chat inside that project. Think of it as a study room.

When Projects beat Skills

Use a Project for context that belongs to one initiative. Use a Skill when several Projects need the same procedure. If you find yourself copying the same instructions into multiple Projects, that is the moment to extract a Skill.

Pricing tier: Claude Pro and above. Not available on the free plan.

Subagents: the specialized worker

A subagent is Claude with a narrow job description, an isolated context window, and a curated tool list. It runs in its own loop and hands the result back to the parent agent. Anthropic's research team measured a >90% lift on complex tasks from a Lead Researcher plus specialized subagents versus single-agent chat.

---
name: code-reviewer
description: Use after edits to find bad patterns in changed files.
model: sonnet
tools: [Read, Grep, Glob]
disallowedTools: [Edit, Write]
---

You are a code reviewer. Find bad coding patterns.
Output: a list of {file, line, issue, suggested-fix}.

The catch: subagents cost roughly 15× more tokens than a single-agent chat. Route the orchestrator to Opus, the specialists to Sonnet, and the read-only file scouts to Haiku, or the bill stops being fun.

Pricing tier: Claude Code and the Agent SDK. Custom subagents live in ~/.claude/agents/ or the project's .claude/agents/.

MCP: the universal plug

MCP — Model Context Protocol — is the wire format that lets Claude talk to an outside system. A Slack MCP server lets Claude post in channels. A Postgres MCP server lets Claude run SQL. MCP is the only tool on this list that gives Claude live, two-way access to systems that change while you are chatting.

MCP is also the tool teams over-use. Every tool you wire in eats from the same context window, and accuracy degrades past about 30–40 connected tools. The fix is not more MCPs — it is fewer, better ones, plus Skills that orchestrate them.

Pricing tier: Claude Desktop, Claude Code, and the API. Browse 3,200+ servers in the MCPFlix marketplace.

Pricing tiers at a glance

Free

Prompts + Skills

Every Claude user

Pro

+ Projects

200K context, RAG

Code / SDK

+ Subagents

Isolated context, hooks

Desktop / API

+ MCP

Live external systems

A 30-second decision tree

  1. Is this a one-off? Use a prompt. Stop reading.
  2. Have you typed nearly the same thing twice this week? Make a Skill.
  3. Is the work organized around a body of background documents? Open a Project.
  4. Does the task have a narrow scope and benefit from its own tool budget? Spawn a subagent.
  5. Do you need live data from an outside system? Wire in an MCP.

Most teams need all five eventually. The lesson is just to stop using one of them for everything.

How MCPFlix fits

The Stack Planneroutputs all five recommendations — not just the MCP list. Tell it what you are building and it returns: "use a Skill for X, a Project for Y, these three MCPs for Z." The Skill Builder gives you a SKILL.md file you can drop into ~/.claude/skills/.

Want the design rules for a good SKILL.md? Read the 6 Skill design principles. Already drowning in plugins and MCPs that do nothing? Clean up your Claude Code setup first.

References