"AI will automatically research, write articles, sort emails, and finish reports." Every time people hear this, they close the tab thinking, "Amazing, but impossible for a non-engineer like me."
This is a lie. You can create your own "private AI agent" today using only Claude's standard features, without code, API keys, or terminals. It takes about 15 minutes per agent once you're used to it.
My name is tatsuki (@nobel_824). I support AI utilization for small and medium-sized enterprises, helping them implement Claude and Codex into their operations. As someone who runs Claude Code all day, I can tell you that too many people overcomplicate "agents." The reality is much simpler and more practical.
This article is based on Khairallah AL-Awady's guide, which was read 2.98 million times overseas, restructured for Japanese users and side-hustle beginners. By the time you finish reading, you will have three types of agents and copy-pasteable design prompts.
Distributing 300 Selected Prompts for Claude
I've made these available at the end of this article.
What is an AI Agent, in Plain English?
Forget all technical definitions for a moment.
An AI agent is simply a state where Claude automatically executes multiple tasks in sequence that you used to do one by one. Nothing more, nothing less.
Think about how you usually use Claude. You ask it to "research this theme." You read the result. You ask it to "create an outline." You check it. You ask it to "write the body." You fix it. You ask it to "format it." You copy it. That's 4 steps. Each time, you are reading, approving, and typing the next instruction.
An agent does these four all at once. If you say, "Research this theme, create an outline, write the body, and format it for publication," Claude handles everything in order and gives you only the finished product.
That's the only difference. A chatbot does one thing and stops. An agent performs a series of tasks and delivers the result. Once you grasp this, you just need to decide which tasks to delegate.
Three Types of Agents You Can Build Without Code
You don't need APIs or agent SDKs. You can create the following three types using only Claude's existing features without writing a single line of code.
- Type 1: Chat-based Agent (Claude Projects) — Write the procedure in the system prompt, and it automatically executes all steps when given a theme.
- Type 2: File Processing Agent (Claude Cowork) — Reads folders on your PC, processes files one by one, and organizes the output.
- Type 3: Scheduled Agent (Cowork + /schedule) — Wakes up at a set time, performs work, and saves the results.
Choose based on what you want to do. Type 1 is for "one-off finished products," Type 2 is for "clearing out accumulated files," and Type 3 is for "tasks you want done automatically every day."
One prerequisite: Type 1 (Projects) can be made on the free plan, but Claude Cowork used in Types 2 and 3 is exclusive to paid plans (Pro / Max, etc.). Furthermore, the /schedule feature in Type 3 is being rolled out gradually, so it may not be available in your environment yet. Let's start by making one Type 1 agent, which costs nothing.
Type 1: Research → Article Writing Agent (15 minutes)
This is a chat-based agent that returns a finished article when given a single theme. One input, one output.
The process is simple. Open Claude, create a new project from "Projects" on the left, and name it something like "Article Agent." Then, just paste the following design prompt into the project's "Custom Instructions (System Prompt)."
You are an autonomous article production agent. When the user provides a theme, execute the following steps automatically without asking for approval in between.
[1 Research] Search for the latest information on the theme on the Web, and extract key points, numbers, and expert views from 5-7 sources. [2 Angle] Identify what readers currently believe about the theme and choose an angle that either overturns or expands that belief. Choose the angle that will get the most engagement. [3 Outline] Create a detailed outline including an introduction starting with the "gap between common sense and reality," 5-7 bold subheadings, specific numbers or examples in each section, and a closing CTA. [4 Writing] Write the body in 2000-3000 words. Max 3 sentences per paragraph. Bold the most important point in each section. Use specific numbers rather than vague expressions. Conversational tone, zero fluff. [5 Review] Self-check based on the following criteria: Does each section add new information? Does every claim have a specific number or example? Are there any parts where a busy person would stop reading (fix if so)? Is the introduction truly compelling?
Finally, output only the finished article with a recommended title at the beginning. Execute steps 1-5 in a single reply without seeking approval.
Now, just open a new chat within the project and type a theme like "AI Agents for Beginners." Claude will run from research to self-review in one go, returning a finished article from a 3-word input.
The key here is the granularity of the instructions. Even for "write an article," the results differ completely based on granularity.
- ✗ "Write an article about AI"
- ◯ "Research → Outline → Body → Review all at once for an article about the first agent an AI beginner should build. The readers are side-hustle beginners who are non-engineers."
Agentization is essentially fixing these "◯" instructions into the prompt instead of typing them every time. Once you understand this, your perspective shifts to "I can do that task too."
Type 2: Folder Batch Processing Agent (15 minutes)
Next is an agent that cleans up files on your computer. From here, we use Claude Cowork (desktop app, paid plan).
If you pass a large production folder immediately, it becomes heavy to redo if it fails midway. I recommend starting with a small folder containing 2-3 files before switching to the real one. The "instruction granularity" applies here too.
- ✗ "Summarize the PDFs in this folder"
- ◯ "Read all PDFs in /Downloads, organize each file into 5 summary points + 3 action items, and save as a .md file with the same name in /Summaries. Finally, create an all-summaries.md that summarizes everything."
Paste the "◯" granularity directly into Cowork. Open Cowork, grant access to the folder you want to process, and ask for batch processing like this:
Look at the /Downloads folder and do the following for all PDFs inside: 1. Read the body 2. Create a summary (up to 5 bullet points) 3. Identify 3 important action items 4. Save a summary file with the same filename and .md extension in /Summaries When all files are processed, create "all-summaries.md" in chronological order. Do not stop for each file; process until the end.
Claude will read the PDFs one by one, create summaries, output individual files, and finally compile a master document. If you have 20 PDFs, hours of manual reading and note-taking are finished with one instruction.
In my business support, this "boring" batch processing is often what people find most effective. Rather than flashy AI, a presence that quietly cleans up accumulated documents is more appreciated on the front lines.
Type 3: Scheduled Agent That Works While You Sleep (15 minutes)
The third is an agent that runs automatically every morning. This is the true power of agents; once set, it keeps moving "even if you do nothing."
There is one thing to do first. Since this example reads Gmail and Google Calendar, you must first enable both connections (connectors) in the Cowork settings. If these are not connected, it will stop with "cannot read email" even if you paste the prompt. This is where I stumbled at first, so clear this hurdle beforehand.
Once connected, type /schedule in the Cowork chat input. The skill to create periodic tasks will launch, so choose the frequency (daily, weekdays, weekly, hourly, etc.) and write what to do once. Set an instruction like this:
Execute every weekday morning at 7:00 AM. 1. Check emails received in Gmail after 5:00 PM yesterday. 2. Categorize each email into "Action Required," "For Reference," or "Ignore (Newsletters, etc.)." 3. Draft reply proposals for "Action Required." 4. Check today's Google Calendar schedule and note participants and agendas for each meeting. 5. Save everything as "morning-brief-[today's date].md" to Google Drive. Use headings in the order of "## Action Required Emails," "## Today's Schedule," and "## Reference."
If you set the frequency to "Weekdays at 7:00 AM," you will wake up every morning with emails sorted, reply drafts ready, and your schedule summarized.
Here is a tip for "working while you sleep." Scheduled tasks run remotely even if your PC is asleep or the app is closed, provided the storage and data are entirely on the cloud side (Google Drive, email, etc.). Conversely, if reading/writing to local folders on your PC is involved, it only runs while the PC is on. That's why in the example above, I intentionally set the storage to Google Drive instead of local. If you want it to move while you are literally sleeping, the trick is to keep both input and output on the cloud side.
One more note: the UI display and /schedule behavior may change with updates. If you can't find it as described, follow the latest guidance within Cowork. The design philosophy of "write once, run automatically" is more essential than the specific screen layout.
3 Habits to "Grow" Your Created Agents
Your first agent will usually be "okay." It won't be perfect. This is where the real work begins; the difference lies in how you grow it.
First is Correction Rules. If you want to fix an output, don't just fix it and move on; add it as a new rule to the system prompt. If you feel the summary is too long, add a line saying "Each summary must be under 100 characters." Repeat this 10 times, and the agent becomes remarkably precise.
Second is Example Rules. Upload several examples of "ideal output" to the project's knowledge and instruct it to "match this standard." An agent with examples clearly outperforms one with only instructions.
Third is Weekly Review. Once a week, look over the outputs from the last 7 days and reflect the good and bad points into the instructions. The completion level after one month will be completely different between an agent with this humble habit and one that is just left as-is.
In other words, an agent is not "done once made," but a partner that gets smarter the more you grow it. I also add my frustrations to my own agents' prompts whenever I notice them.
5 Agents You Can Build Today
Once you understand the 3 types, the applications are infinite. Here are 5 you can make right away:
- Content Repurposing: Give it one long article, and it generates an X thread, LINE message, Instagram caption, and newsletter intro all at once.
- Competitor Watch: Every week, search for 3 competitors and summarize new products, price changes, and shifts in messaging into a brief.
- Customer Onboarding: Give it a customer name and project summary, and it prepares a welcome email, progress schedule, and interview items with the customer's name.
- Meeting Minute Formatting: Give it scribbled notes, and it formats them into decisions, actions with assignees, and next steps for sharing.
- Receipt Processing: Scan the /Receipts folder monthly to extract date, store name, amount, and category into a summary table.
All of these are tasks with "multi-step manual work." Think of one task you repeat every week. That is the next agent you should build.
Summary
- An AI agent is a state where Claude automatically executes tasks in sequence that you used to do one by one.
- There are 3 types you can build without code: Chat-based in Projects, File Processing in Cowork, and Scheduled with /schedule.
- The core of a design prompt is to "write the steps in order and tell it not to ask for approval."
- Agents are not finished once made; they get smarter as you grow them with correction rules, examples, and weekly reviews.
- Once you build your first one, you'll start seeing all repetitive tasks around you as things that can be "agentized."
If you read this and just think "I see," you'll end up doing everything manually again tomorrow. Most people do. Will you spend 15 minutes once to build one, or keep doing it by hand? This is the turning point.
Start by making just Type 1 today. Create a project in Claude and paste the prompt above. That will be your "first one."
I am distributing "300 Selected Prompts for Claude" for free on LINE, which summarizes prompts like the ones used in this article into a copy-pasteable format for various purposes.
▼ Register on LINE to receive "300 Selected Prompts for Claude"
https://t.co/90omRA4UQ7
Please join the Open Chat, and then from the official LINE mentioned in the notes,
send the word
Prompt
to receive the benefits.
Please refrain from speaking within the Open Chat.
