With MCP, an AI agent can reach your tools. But having a kitchen full of ingredients doesn't make you a cook: you also need recipes. Agent Skills are those recipes: small folders of know-how that teach an agent how to do a specific job well, loaded only when the job comes up.
The recipe box story
Your family recipe box has dozens of cards. You don't read them all before cooking. You glance at the tabs (“Biryani”, “Dosa”, “Kheer”), pull out the one you need, and follow it step by step. That's exactly how an agent uses skills. Try it:
📚 Watch an agent pick a skill
Give the agent a task:
The agent's skill shelf
exam-study-plan
Builds a day-by-day revision plan before an exam.
resume-review
Reviews a fresher's resume for SDE roles and suggests fixes.
pdf-forms
Fills in PDF forms from details the user gives.
brand-slides
Makes slides in the college club's colours and fonts.
The agent always carries just the short descriptions (about 400 words for 4 skills). Loading every skill in full on every message would cost about 8,200, and that gap grows with every skill you add. Numbers are illustrative.
What's inside a skill
exam-study-plan/
├── SKILL.md ← required: name, description, then the steps
├── scripts/ ← optional: code the agent can run
│ └── make_plan.py
└── references/ ← optional: longer docs, opened only if neededHere's the whole SKILL.md for a study-plan skill. The part between the --- lines is the header:
---
name: exam-study-plan
description: Builds a day-by-day revision plan before an exam. Use when a student asks for a study plan, revision timetable or exam schedule.
---
# Exam study plan
## Steps
1. Ask for the exam date and the list of topics (if not given).
2. Run `scripts/make_plan.py` with the days left and the topics.
3. Show the plan as a table. Keep the last day for revision only.
## Rules
- Hardest topics go first, while energy is high.
- Never plan more than 3 topics in one day.And the script it runs. We ran it for 4 days and 5 topics; the output below is real:
# make_plan.py: spread topics over the days left, keep the last day for revision
import sys
def make_plan(days, topics):
study_days = days - 1 # last day = revision only
plan = {d: [] for d in range(1, study_days + 1)}
for i, topic in enumerate(topics): # round-robin, hardest first
plan[i % study_days + 1].append(topic)
plan[days] = ["Revise everything + flashcards"]
return plan
if __name__ == "__main__":
days = int(sys.argv[1])
topics = sys.argv[2].split(",")
for day, items in make_plan(days, topics).items():
print(f"Day {day}: {', '.join(items)}")Day 1: DP, Sorting
Day 2: Graphs, Hashing
Day 3: Trees
Day 4: Revise everything + flashcardsProgressive disclosure: three levels
- Always loaded: each skill's name and description (a line or two).
- When a task matches: the full SKILL.md instructions.
- Only if needed: extra reference files, and scripts. Scripts are run, so only their output enters the AI's memory, not the code.
That's why an agent can have dozens of skills and still stay fast and focused.
Skills vs MCP vs prompts vs RAG
| Piece | Gives the AI… | Example |
|---|---|---|
| Prompt | One-off instructions, right now | “Explain recursion with a cricket example.” |
| Skill | Reusable know-how for a kind of task | “How we write lab reports”, with a script |
| MCP | Access to tools and live data | Read Drive, create GitHub issues |
| RAG | The right facts from your documents | Answer from the 50-page handbook |
They stack together. A modern agent might use a skill for “how to prepare a weekly report”, which tells it to fetch numbers through an MCP server, look up the policy with RAG, and follow your prompt for the tone.
🔍 Use a skill when you see
- ✓You repeat the same kind of task often
- ✓There's a right way to do it: steps, format, house style
- ✓It needs a script, template or reference doc alongside
- ✓You want it to work across agents that support SKILL.md
🚫 Don't use it when
- ✗A one-off question (just prompt)
- ✗The AI needs access to a service or live data (MCP)
- ✗It needs facts from a big set of documents (RAG)
⚡ Memory tricks
- 1
“MCP = hands, Skills = how-to”
Access to tools vs know-how for a task.
- 2
“Read the tabs, open one card”
Descriptions always loaded; full skill only when a task matches.
- 3
“Name + description, then the steps”
The shape of every SKILL.md.
- 4
“Say WHAT and WHEN”
A good description says what the skill does and when to use it.
Where skills came from
Anthropic introduced Agent Skills and then released the format as an open standard, documented at agentskills.io. Many AI agents and coding tools now read the same SKILL.md format, so a skill you write once can travel with you. As with MCP servers, only install skills you trust: they can include scripts that run on your computer.
🎯 Spot the pattern: is it a skill?
In an interview nobody tells you the pattern. Read each problem and decide. Look for the signal words.
1. Teach the agent your team's exact steps for writing release notes, every time.
2. Let the agent read and send emails from your Gmail.
3. Package a Python script that fills PDF forms, with instructions on when to run it.
4. Ask a one-off question: “Explain recursion with a joke.”
5. Make slides that always use your college fest's colours, fonts and logo.
🧠 PATH memory card
Remember Agent Skills for good
Picture it · Anchor it · Trace it · Hit-recall it. Four quick steps that turn “I read it once” into “I can explain it in an exam”.
- P
Picture it
Your brain remembers scenes far better than definitions.
← swipe to see the whole picture →
A recipe box: you read the tabs, pull out only the biryani card, and follow it fully. Agents do the same with skills.
- A
Anchor it
A silly sentence you can't forget, where every word maps to the concept.
“Recipe box: read the tabs, open one card, follow it fully”
- Recipe box→The agent's collection of skills (folders)
- Read the tabs→Only names and descriptions are always loaded
- Open one card→Load the full SKILL.md for the matching task
- Follow it fully→Do the steps, run the scripts, respect the rules
- T
Trace it
Run it by hand on a tiny example once. Doing beats reading.
“5-day DBMS study plan” → descriptions scanned →exam-study-planmatches → SKILL.md opened →make_plan.pyrun → plan shown as a table. The other skills stay closed. - H
Hit-recall it
Test yourself now, then again on the day each card comes back.
Card 1 / 6 · new0 mastered · 0 practisedSay the answer out loud first, then flip and be honest.
Where to go next
You now know the modern agent toolkit: a model that predicts words, facts from RAG, actions through MCP, and know-how from skills. Next, see how AI makes pictures from words in how AI image generators work, and then how to stay safe from the scams these tools make possible.
Key takeaways
- ✓An Agent Skill is a folder of know-how: SKILL.md plus optional scripts and references.
- ✓SKILL.md starts with a name and description, then plain-language steps.
- ✓Agents load only descriptions until a task matches (progressive disclosure).
- ✓Skills = how-to; MCP = access; RAG = facts; prompts = one-off instructions.
- ✓Skills are an open standard; install only ones you trust.
Frequently asked questions
What are Agent Skills in simple words?
An Agent Skill is a folder of know-how for an AI agent: a SKILL.md file with instructions for one kind of task, plus optional scripts and reference files. The agent sees only each skill's name and short description until a task needs it, then opens that one skill and follows it.
What goes inside a SKILL.md file?
A short header (YAML front matter) with two required fields, name and description, followed by plain Markdown instructions: the steps, rules and examples for the task. The description matters most: it's how the agent decides when to use the skill.
What is the difference between Agent Skills and MCP?
MCP gives an AI access: a connection to tools and data such as GitHub or a database. Skills give it know-how: how to do a particular job well. A skill can even tell the agent which MCP tools to use and in what order. They're complementary, not competitors.
Why not just put everything in one long prompt?
Because the AI's working memory (its context window) is limited and costs money. With skills, the agent carries only the short descriptions and loads the full instructions for the one skill a task needs. This is called progressive disclosure, and it lets an agent have dozens of skills without slowing down.
Who created Agent Skills, and which tools support them?
Anthropic created the format and released it as an open standard (documented at agentskills.io). Many AI agents and coding tools now support the same SKILL.md format, so a skill you write can work across different agents.
Are skills safe to install?
Treat a skill like any software: skills can include scripts that run on your computer. Install skills only from sources you trust, read the SKILL.md before using it, and be careful with skills that fetch content from the internet.
🧠 Quick check: Agent Skills
1. What are the two required fields at the top of a SKILL.md?
2. An agent has 30 skills. What does it keep in memory all the time?
3. You want the AI to follow your club's exact report format every time. Best fit?
4. You want the AI to create issues in GitHub. Best fit?
Checking your access…