How to Set Up Claude Code Features for Real Work

Answering one email with Claude takes a prompt. Answering forty a week takes setup. The six Claude Code features, skills, connectors, plugins, scheduled tasks, hooks and mods, replace the explaining, the copy-pasting, the forgetting and the guessing one piece at a time.

What the Six Claude Code Features Do

The six Claude Code features, skills, connectors, plugins, scheduled tasks, hooks and mods, each fix one specific part of using Anthropic AI assistant for repeating work. A skill stores your process, a connector gives Claude access to another app, a plugin adds a prebuilt role, a scheduled task starts work on time, a hook enforces a checkpoint, and a mod adds your own interface inside Claude Code, Anthropic agentic coding tool that runs in the terminal.

A simple running example shows why each one exists. You run a small furniture business. Every Monday your inbox holds quote requests, delivery questions and sizing advice. Pasting each email into Claude explaining your process every time works, but it does not scale and nothing improves between Monday and Monday.

The features stack in a fixed order, and each layer depends on the earlier ones:

Here is how the six compare at a glance:

Feature	What it is	Problem it solves
Skill	A folder with a SKILL.md process file	Re-explaining the same procedure
Connector	Links Claude to an app, usually via MCP	Copy-pasting between tools
Plugin	Prebuilt pack of skills and connectors	Building every role from zero
Scheduled task	Runs Claude set times	Remembering to start the work
Hook	Automatic rule at a fixed moment	Hoping the model is careful
Mod	Custom panels inside Claude Code	Not knowing if any of it works

Skills: Write the Process Once

A skill is a written process that Claude follows every time, so you stop typing the same instructions. Anthropic introduced the skills concept in October 2025, and it is documented in the agent skills documentation. The format is deliberately plain: a folder containing a SKILL.md file plus any supporting material.

The structure has three parts. At the top sits a short description, for example: "new inquiry. Use this when someone emails asking for a quote." That description is all Claude normally sees, so with fifty skills installed it reads descriptions like a cookbook's contents page and opens only the relevant file. Below the description come the ordered steps: read the email, work out what the customer wants, check the price list, suggest the best option, and ask for anything missing. Then come worked examples, previous replies you were happy with, which show Claude what a good output looks like.

The folder can also hold reference files and small scripts. A script is useful for anything that must be identical every run, such as totalling a quote, because Claude executes it rather than recomputing from scratch. After setup, a new email needs no explanation at all: invoke the skill by name, paste the email, and Claude follows the same steps to the same standard.

Connectors: Give Claude Access to Your Apps

A connector plugs Claude directly into another application so it can read and act there instead of you shuttling text between windows. Under the hood most connectors use the Model Context Protocol, an open standard Anthropic announced in November 2024 and documented at modelcontextprotocol.io. In practice, when someone says MCP in this context, they usually mean a connector.

Setup is an account link, not a project. In Claude connector settings you authorise Gmail, and from then on Claude search the inbox, read messages and write drafts, with sending controlled by which permissions you grant. The Monday run changes shape completely: you ask Claude to check the inbox for new inquiries from the last 24 hours, it reads each email, runs your new-inquiry skill, and places three reply drafts in Gmail for you to review and send.

The division of labour is worth stating plainly, because it is the core pattern of all six Claude Code features. The skill is the how. The connector is the access. Either one alone gets you halfway.

Plugins: Add Roles You Did Not Build

A plugin is a ready-made pack of skills, and often connectors too, assembled for a particular job. It works like hiring someone who already knows the role instead of training from zero. Anthropic maintains an official plugin marketplace for Claude Code, documented at docs.Anthropic.com, which is the safe place to start.

For the inbox example, a customer-support plugin handles returns, warranty questions and complaints, and a sales plugin chases quotes that went quiet. Installation drops those skills alongside your own, so the inbox now has two or three roles working in parallel: one for new inquiries, one for support, optionally one for follow-up.

The catch is the same as with a new hire. The plugin does not know your return policy, your warranty terms or your tone. Open it, read what is inside, and add your specifics: what it may promise, what it must escalate, and how you talk to customers. An unreviewed plugin drafts confidently in a voice you never approved.

Scheduled Tasks: Start the Work on Time

A scheduled task tells Claude when to act: what time, on which days, doing what. Without one you still open Claude every morning and say "check my inbox", and on the days you forget or are busy, the emails just sit there. With one, the work begins whether you remember or not.

Sizing the schedule to the workload matters. A handful of emails a day justifies one morning check; faster response times justify a second run after lunch. In the example, Claude checks the inbox every weekday at 9:00 and 13:00, routes new inquiries through the inquiry skill, routes support email through the support plugin, and leaves anything unmatched for human attention. You open Gmail at nine to drafts already waiting.

One practical constraint from the walkthrough: if the task runs on your machine, the machine must be on at the scheduled time. Claude also run scheduled tasks in the cloud when you need the schedule to hold regardless of whether your computer is awake.

Hooks: A Checkpoint That Never Forgets

A hook is a rule that runs automatically at a specific moment, every time that moment occurs. The distinction from a skill is enforcement. A line in a skill saying "be careful with pricing" is advice Claude tries to follow; a hook attached to the send-email action runs whether Claude remembers or not. Hooks are a documented Claude Code mechanism, described at docs.Anthropic.com.

The example places the hook just before an email is sent. The hook checks the draft: does it mention a price, a discount, a refund or the warranty? If none of those appear, it goes out. If one does, the hook stops the send and asks for approval. Everything else in the workflow keeps running; you are pulled in only for the messages that actually need a human.

Two failure modes from real use are worth recording. First, automation drift: once drafts got good, the speaker let Claude send simple replies alone, and twice it said something wrong, offering a discount the business does not give and misstating warranty coverage. Second, hooks only catch what you list. A hook checking for "refund" will not catch "money back", so the keyword list needs maintenance.

Mods and the Feedback Loop: Measure, Then Improve

A mod changes Claude Code itself, letting you add your own panels and buttons inside the interface. In the walkthrough this becomes a small dashboard next to the chat, and it answers the question automation creates: is any of this working?

The dashboard needs data, so one step joins the skills: every time Claude handles an email it appends one line to a log with who it was, the email type, the quote and the outcome. The mod reads that log and renders the week. In the video's example run that was 12 new inquiries, 10 quotes sent, and 3 accepted, a speaker-reported figure from a small sample. Ten decisions is a hint, not a verdict, but it beats guessing, and it enables real experiments: change the package or the pricing, and next week's dashboard shows the effect. A second weekly task can read support conversations and score how they ended, giving a customer-happiness figure on the same panel.

The last step reuses everything already built. A Friday afternoon scheduled task reviews the week: which questions the skill failed to cover, which replies got a yes, which drafts you rewrote before sending. It writes suggested skill changes, such as asking for the delivery address earlier or dropping a package nobody picks, into an improvements tab. You approve what makes sense and Claude updates the skills. You stay in charge of changes, but you are no longer the only one noticing them.

Which Claude Code Features Should You Set Up First?

Start with a skill, then add each other feature only when you hit the problem it solves. The video's own guidance is that you do not need all six, and for most tasks the sequence below is the cheapest path to a workflow that holds up:

1. Write a skill for the process you repeat most, with a clear description, explicit steps and two good example outputs.
2. Add a connector when copy-pasting between Claude the app becomes the bottleneck.
3. Install a plugin when the task has a role you have not built, then customise it with your own policies.
4. Schedule the task once you trust the output but keep forgetting to start it.
5. Add a hook the first time the automation says something you would not have said.
6. Add a mod when you need to know whether the whole thing is working, not just that it is running.

Reviewed this way, the six Claude Code features are less a checklist than a diagnosis kit: each one maps to a distinct failure you will recognise from your own week, and the setup effort is only justified for the failures you actually have.

Frequently Asked Questions

• What is a Claude skill? A skill is a folder containing a SKILL.md file with a description, ordered steps and example outputs, plus any reference files or scripts. Claude reads the description first and loads the full skill only when it matches the task, so the process is written once and followed consistently.

• Do connectors and MCP mean the same thing? Most connectors use the Model Context Protocol under the hood, which Anthropic released as an open standard in November 2024. In everyday use the terms overlap: a connector is the user-facing link between Claude another app, and MCP is the protocol making it work.

• Are plugins safe to use without editing? No. A plugin arrives without your return policy, warranty terms or tone of voice. Treat installation as hiring, not deployment: read its skills, add your specific rules, and state what it is allowed to promise customers.

• What is the difference between a skill instruction and a hook? An instruction in a skill is advice the model tries to follow; a hook is a rule that runs automatically at a fixed moment, every time. For anything sensitive, such as pricing or refunds in outgoing email, a hook is the correct mechanism because it cannot be forgotten.

• Do scheduled tasks require my computer to be on? If the task runs locally, yes, the machine must be awake at the scheduled time. Claude also supports running scheduled tasks in the cloud when you need the schedule to hold independently of your hardware.

From a Watched Video to a Written Reference

This article exists because the clearest explanations of a workflow often live in a video: the walkthrough above, by Gustavo Dev Doido of Crazystack Typescript and the Bootcamp do Dev Doido community, teaches the six-feature progression far better in Notion than most documentation does in prose. If your own videos hold knowledge like this, the same transformation is available to you.

Skala Blog turns a YouTube video into a structured, source-linked article: paste the URL, the video is transcribed, and you get a draft you review before publishing. The inbox walkthrough you just read is an example of the format. More from this creator and the broader Crazystack catalogue is at crazystack.com.br.

Source video