Learn skills, connect them, then build your own.
Everything in one place, in five steps. Start at the top, or jump straight to the part you need.
Prompts drift. Skills compound.
A prompt is a one-time instruction you retype every session. A skill is a living document that keeps your procedures, decision logic, and corrections in one place, so the work gets more accurate the more you use it. Here's what one actually is.
my-skill/
├── SKILL.md ← instructions (required)
├── preferences.md ← what it learns from you
├── references/ ← supporting docs
│ └── examples.md
├── scripts/ ← executable logic
│ └── calculator.py
└── assets/ ← templates, images
└── template.docxA skill is a folder of intelligence
At its core, a skill is a folder of files that teach AI how to do something specific. One file does most of the work, the SKILL.md:
- Metadata: the name, description, and when AI should use it
- Instructions: the actual procedures and decision logic
- References: links to the other files, including the preferences it learns into
Everything else is optional. You can build a powerful skill with just a well-written SKILL.md. For the format field by field, read the SKILL.md anatomy tutorial.
The three layers inside it
What a skill knows sorts into three layers. Separating them is what lets one skill keep paying off without constant rewrites:
Foundation: portable know-how
The transferable expertise you would bring to any job: triage heuristics, deal-analysis frameworks, coding standards.
Your context
Your VIP senders, brand voice, payer rules, investment thesis. This is the layer the feedback loop keeps refining.
Right now
Active deals, current projects, this week’s priorities. Updated often.
How a skill learns
This is where "compound" comes from. You give the skill one standing instruction to capture your feedback. After that it runs the loop itself every session, with no re-prompting, asking what to change and saving the answer. It isn't the model rewriting itself unprompted; it's a loop you set up once:
- 01
Run the skill
It does the work using its current instructions.
- 02
You correct it
Flag what it got wrong, miscategorized, or should handle differently.
- 03
It saves the lesson
Your corrections are written to the preferences.md file inside the skill.
- 04
Next run is sharper
The skill reads those preferences first, so it starts where you left off.
The loop repeats every run. Prasad's own email-triage preferences file reached 23 revisions over 7 to 8 weeks, and the assistant kept getting better.
That standing instruction is just one line: "After every run, ask me what to change, and update your preferences." No code, no retraining, and nothing changes unless you confirm it.
How skills stack
Skills aren't used one at a time. You can load several complementary skills into a single conversation so they work in concert: the AI applies all of them at once. Stack the right few and a workflow that used to mean switching between tools happens in one session.
- Brand voice
- Image prompt
- Blog generator
- Answer-engine optimization
One published blog post used all four together in a single Claude conversation: how the team writes, its visual style, the content structure, and the optimization that helps it rank in AI-powered search. The same approach works for presentations, sales outreach, or any workflow where consistency matters across a team.
One connection. Every tool you already use.
Skills aren't locked to one app. Store them once in Agentman and reach them from any AI tool through a single MCP connection, the open standard for giving AI access to your tools and content.
- Claude
- ChatGPT
- Cursor
- Your agents
Connect in under five minutes
Listed in Claude’s Connectors Directory
- 01
Register free
Create your workspace at studio.agentman.ai/skills and add the skills you want.
- 02
Connect your AI
In Claude, find Agentman Skills in the Connectors Directory and choose Connect. In other tools, add a connector and paste skills.agentman.ai/mcp.
- 03
Authorize & verify
Approve access, then ask "list my skills" to confirm they're live.
Operator or developer, the steps are identical. You paste one URL: nothing to install, no engineering required.
Open the connect guideGo deeper on the path that fits your role.
Same foundation, different next steps. Each path is a short, ordered sequence that ends by handing you off to the right way to build.
Put skills to work
Automate your own and your team's workflows. No code required.
- FoundationsWhat are agent skills?Prompts vs. skills
- Industry applicationsExplore the Skills LibraryFilter by Healthcare, Marketing & Sales, or whatever you work on.
- Take actionCreate a skill
Build skills into your products
Author robust skills and embed them in your own apps and agents, following the Agent Skills spec.
- FoundationsSKILL.md anatomyWhat makes a great skillExplore the Skills Library
- Build your first skillCreate a skillGet API access
- Go deeperAgentman on GitHub
Watch a self-learning skill get built.
The fastest way to understand skills is to watch one come together. In 30 minutes, build a Gmail assistant that triages your inbox, then wire up the feedback loop from step 01 so it sharpens every time you correct it.
Build a Self-Learning Gmail Assistant
Sort important mail, summarize newsletters, flag spam, plus the feedback loop that writes your corrections back into the skill. No code required.
Hosted by Prasad Thammineni · Beginner-friendly
Open the full lessonCreate your own skill.
You've seen what a skill is, how it learns, and how to connect it. Now make one. Start the fast way, then refine in the visual editor. Your skill works in Claude, ChatGPT, and your agents.
- Recommended · fastest
Quick Create
Describe what you need in plain English. The builder writes the SKILL.md, organizes your content, and formats it correctly, feedback loop included.
"Create an email-triage skill that sorts important mail, summarizes newsletters, flags spam, and after every run asks me what to fix and updates its preferences."
Try Quick Create - Best for: adapting existing expertise
Clone & Customize
Start from a system skill that's close, clone it to your workspace, then adjust the thresholds, templates, and decision logic for your team.
Browse skills to clone - Best for: unique organizational knowledge
Build from Scratch
Open the visual editor with a blank SKILL.md. Full control over structure, content, and organization.
Open the editor
Quick Create: from description to skill in about 2 minutes
- 01
Describe it
Say what the skill does, when AI should use it, and the key decision logic.
- 02
Review the structure
The builder generates a SKILL.md with frontmatter, instructions, and a decision tree.
- 03
Edit & refine
Adjust thresholds, add edge cases, and customize templates in the editor.
- 04
Publish & use
Save to your workspace. It is available in Claude, ChatGPT, and your agents straight away.
The visual skill editor
Where you refine what Quick Create wrote, or start from a blank page. Everything a skill needs is on one screen.
- SKILL.md
- preferences.md
- references/
- scripts/
- assets/
Use when an invoice arrives from a supplier and needs checking before it is paid.
Steps- Match the invoice to its purchase order.
- Compare quantities and unit prices.
- Flag any difference over 2%.
Metadata panel
Name, version, description, category and tags in form fields. No YAML required.
Content editor
Rich text for SKILL.md, with a markdown preview that shows exactly what the AI will see.
File manager
Drag and drop into references/, scripts/ and assets/ to organize supporting content.
Token counter
A running token count for OpenAI, Anthropic and Google models, so you know your context cost.
Version history
See every version, compare two side by side, and restore any one with a click.
Preview mode
Test in a sandbox to see how AI applies your instructions before you publish.
Lessons, working skills, help and tutorials.
Four places to go once the five steps are done.
Lessons
Build a self-learning Gmail assistant in 30 minutes, with the recording, the prompt and the setup steps.
Open the lesson
Skills library
Working skills you can read, run and clone. Filter by the work you do.
Browse the library
Help Center
Step-by-step answers for sharing, exporting and connecting skills, and for building agents.
Open the Help Center
Blog tutorials
Longer guides on writing, structuring and governing skills, tagged Learn.
Browse the tutorials
Tutorials to read next
How to build your first agent skill: SKILL.md anatomy
The SKILL.md format field by field: the two required frontmatter fields, the optional ones, when to add scripts and references, and three ways to build.
What are AI skills? The complete guide
What skills are, how they work, and why teams are moving from ad-hoc prompts to written procedures.
A prompt is a question. A skill is a procedure.
Why the best prompt on your team keeps getting lost, and what changes when the procedure is written down once.
You don’t need to code to build agent skills
The plain-English workflow that takes a non-developer from an idea to a published skill.
Claude Skills vs Agent Skills
The same open SKILL.md standard. The difference is where the skills live and who governs them.
Common questions about agent skills
What is an agent skill?
A written procedure an AI agent follows: the steps, the judgement calls, the standards, and what "done" looks like. Prompts drift between people and between sessions. A skill is versioned and repeatable, which is why skills compound and prompts do not.
How is a skill different from a prompt?
A prompt is a question; a skill is a procedure. Two people asking the same question get different results, and nobody can audit what was asked. A skill runs the same steps every time, can be reviewed and improved, and carries its version history with it.
Do I need to know how to code?
No. Skills are written in natural language. The person best placed to write one is usually the person who already does the work, not an engineer — that is the point.
Which AI tools can use these skills?
The ones you already use. One MCP connection makes your skills available across your tools, so a skill you write once is not locked to the assistant you happened to write it in.
How long does it take to get through this?
The path is five steps: what a skill is, connecting your AI, watching one get built, building your own, and then going deeper on the track that fits your role. You can do the first three in a sitting.
Ready to build your first skill?
Describe what you need, clone a starting point, or open a blank canvas.