Claude Usage Guide for Beginners: Mastering the Three Interfaces

"I pay for Claude every month, but I end up using it just like ChatGPT..."

This isn't Claude's fault; it's just that you don't know about the existence of the "three interfaces."

Have you had these experiences while using Claude?

• Paying the monthly fee but only using it for chat questions
• Hearing names like Code or Cowork but thinking they don't apply to you as a non-engineer
• Noticing the team next to you is clearly faster despite using the same tool, and not knowing why
• Letting six months pass without knowing the "correct way" to use it

By the time you finish reading this article, it will be clear which interface suits you, and your way of using Claude will fundamentally change.

Anthropic officially designed a three-interface structure. An article by overseas AI practitioner Khairallah, an investor and researcher with 45,000 followers, covering all 40 functions, has caused a huge stir with 184K views.

Based on that information, I will explain specifically where Japanese business professionals should start and in what order, with concrete examples.

Two things before we begin:

1. Save this and set aside 30 minutes this week.
2. Share this with others who have the same struggles with Claude.

Original post here:

https://x.com/eng_khairallah1/status/2050869463252455584

Claude's Three Interfaces and Recommended Workflow

Claude is not just one tool. There are three interfaces: Chat, Cowork, and Claude Code.

Here is the conclusion I want to convey first:

Claude Code is the most powerful, but it takes the longest to get used to.

So, there's no need to jump straight into Claude Code. It's most efficient to step through the experience in stages:

Chat → Cowork → Claude Code

By following this order, you can step up while feeling the benefits of each.

In Chat, you get the sense of "leaving it to the AI." In Cowork, you experience "AI agents moving autonomously." Then, when you move to Claude Code, you'll physically understand why it's the strongest.

Of course, if you can start with Claude Code right away, feel free to do so. But for most people, taking it step-by-step helps them correctly understand the value of Claude Code.

From here on, I will explain in the order of this recommended workflow.

Chat: The Starting Point for Everyone

Positioning:

The most accessible interface. You can start using it as an extension of ChatGPT.

However, if you only "ask questions in chat," you're only drawing out 20% of Chat's capabilities.

Chat alone can sufficiently change your work, but there is a limit. When you feel that limit, it's time to move to Cowork.

■ Teach work context just once

Are you copy-pasting "I'm in the marketing department, in charge of a B2B SaaS company..." every time?

With the Projects feature, if you set the work context once, it is held permanently. Team rules, project background, reference materials—if you register these first, the stress of re-explaining from zero every time disappears.

Example: An English sales team leader puts "Company product overview," "Target customer attributes," and "Past closing patterns" into Projects. Then, brainstorming proposals or competitor comparisons can start directly from the main topic.

■ Let it remember your preferences to make it "yours"

With the Memory feature, you can let Claude remember your preferences.

"Give it to me in bullet points," "Use a casual tone," "State the conclusion first." These types of instructions you used to type every time become unnecessary once taught. It remembers across conversations, so any chat you start is customized to you from the beginning.

Furthermore, with the Styles feature, you can save response tone presets for proposals, emails, or reports. Just switch according to your purpose.

Example: A marketing person creates two Styles: "Internal reports are data and conclusion-first" and "For clients, use polite language + specific numbers." Switching them changes the writing style instantly.

■ Delegate entire research tasks

The Deep Research feature is like hiring a junior analyst. It reads across dozens of sources and summarizes them into a comprehensive report. Research that takes 2 hours finishes in minutes.

Example: Market research for a new business. Just by instructing it to "Summarize the market size, major players, and recent trends of domestic remote work support SaaS," a report comes out. Your job changes from "researching" to "making decisions based on research results."

■ Ask questions to a 50-page report

PDFs, spreadsheets, CSVs. You can drop files directly into the chat. You don't need to read a 50-page report from the start. You can ask directly, "What are the main points on page 37 in 3 lines?" or "Which department has declining sales compared to last year?" You can grasp the gist of materials that used to take an hour to read in 30 seconds.

Example: Upload quarterly financial reports and ask, "Compared to the previous quarter, tell me the most improved and most worsened indicators." Preparation for a meeting that took 30 minutes now ends in 5.

■ Compare multiple options to decide

With the Branching feature, you can create multiple branches from a single message. Have you ever said "Proceed with plan A" and then thought "Wait, what if we go with plan B?" Just by editing the message, a different branch is born. You can look at Plan A and Plan B side-by-side and choose for yourself.

Example: Next term's marketing budget allocation. Simultaneously develop a "70% focus on SNS ads" plan and a "Diversify into exhibitions + content" plan to compare the merits and risks of each. You can consider both possibilities in one conversation.

■ Refine deliverables separately from the conversation

The Artifacts feature separates the dialogue and the deliverable into two layers. Normally, deliverables get buried in the flow of chat. With Artifacts, the deliverable is always displayed in a separate panel. You can write a report while simultaneously consulting on the structure. This "dual-surface work" is possible.

Example: While displaying a proposal draft in Artifacts, you can refine it through dialogue, saying "I want to add more quantitative data" or "Emphasize the differentiation points from competitors." Since you can give correction instructions while looking at the final form, the finish is fast.

Cowork: Experiencing AI Agents

Positioning:

A place to safely experience "what it means to leave work to AI" before moving to Claude Code.

Since it runs in a sandbox, there's no damage even if it fails. Once you realize the power of agents here, moving to Claude Code becomes smooth.

The difference between Chat and Cowork isn't the "number of features." The "way you hand over work" changes. In Chat, you instruct steps one by one. In Cowork, you hand it over and leave it, saying "Prepare this deliverable in this state." From step instructions to result definitions—this shift is the essence of Cowork.

■ Fully automate every morning's routine

With the /schedule function, you can schedule and fully automate repetitive tasks. A review report of last week is automatically ready on Monday morning. Competitor movements are summarized on Wednesday. A draft of the weekly report is prepared on Friday. After setting it up, humans don't need to lift a finger.

Example: Set it once to "Every morning at 9 AM, cross-reference the summary of emails received yesterday with today's calendar and create a prioritized To-Do list." From the next morning, the moment you sit at your desk, your tasks for the day are organized. Your job is to "decide what to automate." Execution is left to Claude.

■ Chain multiple tools with one command

Connector Chaining. Check Gmail → Check Calendar → Get Drive materials → Create meeting prep document. You complete work that spans four tools with a single instruction. Opening 5 tabs, copying, switching, pasting—that time is all "labor," not "judgment."

Example: "Prepare for tomorrow's client meeting. Check participants from the calendar, pull previous agendas from past emails, find the latest proposal materials from Drive, and create a briefing memo." Everything finishes with one instruction.

■ Finish 30 minutes of work in 6 minutes

There is a feature that automatically splits large tasks into 4-5 agents and processes them in parallel. Work that takes 30 minutes if done alone takes 6 minutes if divided into five. It's the same principle.

Example: "Research the latest news of 5 competing companies, summarize each in 200 words, and compile them into a comparison report." Since the research for 5 companies runs in parallel, the waiting time is almost zero.

■ Install plugins specialized for your job

Plugin Marketplace. For PMs, marketing, legal, finance. You can add plugins that suit your profession. Furthermore, with Plugin Chaining, you can link multiple plugins to complete "Research → Analysis → Report" with one command.

Example: For a marketing person, install "SEO analysis plugin + writing plugin + data visualization plugin." "Analyze last month's blog post performance and provide 3 plans for next month with improvement suggestions" works in one shot.

■ Automatically switch rules by folder

Opening Client A's folder automatically applies rules for Client A. Opening Client B's folder switches them.

Example: If a salesperson is in charge of 10 companies, they create 10 folders. In each, they write "Proposals to this company use a casual tone" or "This company requires formal + quantitative data." Just by opening the folder, the mode switches automatically.

Once you experience this "Cowork moving autonomously," you'll feel the urge to "teach it more structurally" or "control it more precisely." When that urge arises, it's time to move to Claude Code.

Claude Code: Strongest but Requires Getting Used To

Positioning:

The most powerful of the three. And Claude Code shows its true potential in tasks other than coding.

However, since it assumes terminal operation, it takes the longest to get used to. Once you're used to it, it has the highest ROI. The decisive difference from other interfaces is that "quality is reproduced permanently with one setting."

*Recently, the barrier has lowered significantly as Claude Code can also be run as a desktop app.

This is the biggest point of misunderstanding. Claude Code is not a "place to program." It is a "place to structurally teach AI your way of working and make it reproducible."

The essence of Claude Code is that it has a "memory" called a file system. Because it has memory, it can maintain context, and because there is context, the quality of judgment improves. Chat and Cowork do not have this "permanent memory and structure."

■ Save work patterns as "Skills"

This is the absolute essence of Claude Code. Skills are definitions of reusable work patterns. "I'm doing the same work every time, but I'm rewriting the prompt from scratch every time." This waste is fundamentally resolved.

Once defined, Skills are reproduced with the same quality no matter how many times they are executed. Prompts disappear when the conversation ends. But Skills remain. Define it in 10 minutes and permanently save hours of work. This is the fundamental difference between Claude Code and other AIs.

Let's see how to use it in Japanese workplaces by profession.

Sales Team: "Meeting Notes → Minutes + Next Actions" Skill

Just by passing notes after a meeting to the Skill, formatted minutes come out. Furthermore, "Action items for next time," "Client concerns," and "Status summary for SFA input" are automatically extracted. You type in notes for 5 minutes while the excitement of the meeting is still fresh. Then the Skill structures it. The work of remembering "what was that talk about" two hours later disappears.

Marketing: "Competitor Press Release → Impact Analysis + Response Plan" Skill

When a competitor issues a press release, just by passing the URL, "Impact on our company," "Response plan when asked by customers," and "Points to counter in the next campaign" are organized. When asked "Company X made this announcement last week" at a Monday morning meeting, you no longer end with "I'll check." You can share the analysis on the spot.

HR/Recruitment: "Interview Notes → Evaluation Sheet" Skill

Inputting scribbled notes written by an interviewer results in an evaluation sheet aligned with recruitment criteria. It is organized with evidence, such as "Communication skills: Answered △△ to question ○○. Corresponds to Criterion B." Variations in writing styles among interviewers and subjective "somehow good" feelings are structured and become comparable. The quality of recruitment meetings changes.

Accounting/Back Office: "Expense Reimbursement Email → Journal Entry Draft" Skill

Expense reimbursement application emails from employees. It reads the amount and purpose, tentatively assigns account titles, and even prepares the addition lines for the monthly report. While the approval flow itself is judged by humans, "input work" and "journal entry judgment" are separated. You can create a state where you can focus only on judgment.

Planning/Business Development: "Market Data → Scenario Analysis" Skill

Inputting market research data or articles organizes them into three plans: "Bullish scenario," "Realistic scenario," and "Risk scenario." Each scenario includes "Prerequisites," "Assumed numbers," and "Judgment points." You can prepare to answer quantitatively when asked "What happens in the worst case?" at a management meeting.

What's common to all professions is that Skills are designed so that judgment remains with humans, while structuring and formatting are handled by the AI. The role of Skills is not to dump work on AI, but to arrange information in a way that makes it easy for humans to judge.

"So, what do you actually write in Skills?"

You're curious, right? It's faster to see the real thing. For example, the content of the sales team's "Meeting Notes → Minutes" Skill looks like this:

1# Organizing Meeting Minutes
2## Task
3Receive meeting notes and organize them in the following format.
4## Output Format
5**■ Minutes**
6- Date/Participants/Agenda
7- Key points of discussion (3-5 bullet points)
8- Client reaction/temperature
9**■ Next Actions**
10- Our side's To-Do (with person in charge + deadline)
11- Items to confirm with the client
12**■ Notes for SFA Update**
13- Presence of phase change
14- Change in closing probability (Up/No change/Down + reason)
15- Expected closing timing
16## Rules
17- For ambiguous parts in notes, write "To be confirmed" (do not fill with speculation)
18- If money is mentioned, always include it in Next Actions
19- If competitors are mentioned, add a separate "Competitor Information" section

That's it. There is no programming at all. You are just writing in Japanese, "Organize in this format" and "Follow these rules." It's just like writing down your way of working to hand over to a junior.

Let's look at one more. The marketing person's "Competitor Analysis" Skill.

1# Competitor Press Release Analysis
2## Task
3Receive competitor press releases or news articles and analyze the impact on our company.
4## Output Format
5**■ 3-Line Summary**
6(Summary that can be reported to a boss in 30 seconds)
7**■ Impact on Our Company**
8- High/Medium/Low and reason for judgment
9- Our products/services affected
10**■ Customer Response**
11- Response plan when asked by customers (1-2 sentences)
12- Items to share with the sales team
13**■ Counter-plan**
14- Our strengths to emphasize in the next campaign
15- Direction of counter-measures (3 plans)
16## Rules
17- Analyze based on facts rather than emotional evaluations ("threat," "no problem")
18- If impact is "High," put items requiring response within 24 hours at the beginning
19- Leave unknowns as "Requires additional research" (do not conclude with speculation)

As you can see, what's written in Skills is a "template for how to proceed with work." What input to receive, what format to output in, and what to be careful about. You are just verbalizing and writing down what you usually do in your head.

Once created, it works with the same quality next week and next month. Quality doesn't drop even if the person in charge changes. This is the meaning of "eliminating dependency on specific individuals."

■ Hand the "Work Textbook" to AI

There is a mechanism called CLAUDE.md. You just write "your judgment criteria," "team rules," and "project regulations" in a text file. Once written, Claude Code continues to move according to it. You no longer need to give verbal instructions every time.

Example: A brand manager writes "Brand Guidelines," "Past NG Expression List," and "Specific Examples of Tone & Manner" in CLAUDE.md. Then, all documents generated by Claude Code automatically come out in a state compliant with the brand guide. The number of times work is sent back in review drops drastically.

■ Make it show the plan before execution

With the Plan Mode feature, you can instruct Claude Code, "Don't just do it. Show me the plan first." You move to execution only after confirming the plan and saying "OK." If there's a sense of discomfort, you instruct corrections. Humans always hold the judgment, and only execution is left to the AI.

Example: When you instruct "Proceed with next month's event planning," Claude Code first presents steps like "Venue booking → Speaker approach → Announcement creation → Customer attraction measures." If you change the priority to "Do the customer attraction measures first," it moves accordingly.

■ Don't let the author review it

There is a mechanism to not let the same session review what Claude Code output. It's hard to notice mistakes if you check a report you wrote immediately after writing it, right? AI is the same. Things generated in the same context become lenient toward their own output. So, you check with a different "eye" in a different session.

Example: Have a different session review the results output by the "Minutes → Action Item Extraction" Skill for "omissions or whether deadline settings are realistic." By separating generation and verification, quality stabilizes.

■ Choose the model according to the task

You don't need to use the highest-grade model for all work. Simple text formatting is enough with Haiku. Use Opus only in situations requiring complex decision-making. It is a human's job to judge "what level of intelligence is required for this task?"

Example: Use Haiku to keep costs down for daily routine report generation, and use Opus for brainstorming next term's business strategy. Just this distinction can sometimes more than halve monthly costs.

Claude Code certainly has a high initial hurdle. You need to open a terminal and be conscious of file structures. But beyond that hurdle is "work quality that is permanently reproduced once set." What you need to re-instruct every time in Chat or Cowork only needs to be done once in Claude Code. When this difference accumulates, the ROI is overwhelming.

Roadmap for the Starting Method That Suits You

"So, what should I start with?"

The answer is simple. The starting point changes depending on the task you feel the most stress with in your current work.

■ If "re-explaining the same thing every time is a hassle" → Start from Chat

10 minutes today: Set up Projects in Chat. Just paste your work context as text. The quality of dialogue from tomorrow will change.

30 minutes this week: Try executing one search with Deep Research. Try leaving one of the research tasks you usually spend 2 hours on to Claude.

Next week: Register your preferences in Memory and preset your writing style with Styles.

■ If "repetitive work is taking up time" → Start from Cowork

15 minutes today: Try instructing "Check Gmail and create today's task list" in Cowork. First, experience an AI agent moving autonomously once.

This week: Automate one morning routine with /schedule.

Next week: Try multi-tool integration with Connector Chaining. Experience moving two or more tools with one command.

■ If "you don't want work quality to depend on specific individuals" → Start from Claude Code

30 minutes today: Write down just one "judgment criterion for your work" in CLAUDE.md. For example, "Structure rules when writing a report."

This week: Define one Skill. Patternize the one that takes the most time among the repetitive tasks you do every week.

Next week: Experience the "Plan → Approval → Execution" flow in Plan Mode. Grasp the sense of division of labor where humans judge and AI executes.

You don't need to start everything at once. It's enough to move to the next after feeling "this is useful" with one interface. That feeling will be the best motivation to move to the next step.

Features Not Yet Introduced Today

It's also useful to grasp the overall picture of features not covered in detail in this article. Please come back when you need them.

■ Chat-related

• Image Analysis: Read and analyze screenshots or photos of whiteboards
• Canvas: Draw flowcharts or organizational charts within the chat

■ Code-related

• /compact: Compress context to maintain continuity when conversations get long
• Multi-File Editing: Edit multiple files at once

■ Cowork-related

• Sandbox Security: Run in a safe sandbox environment. No damage even if it fails
• Browser Bridge: Link web information gathering with local processing
• Token Usage Awareness: Cowork consumes 3-5 times more resources than usual. Optimize with batching

■ Platform Common

• Usage Dashboard: Visualize usage and patterns
• Team Sharing: Share projects, skills, and settings within the team
• Data Privacy Controls: Opt-out of learning use and privacy management

Summary

• Claude's three interfaces. The recommended path is Chat (starting point for everyone) → Cowork (agent experience) → Claude Code (strongest, highest ROI).
• Claude Code is not a tool for coding. It shows its effect in tasks other than coding.
• The reason Claude Code is the strongest is Skills. Once defined, they are reproduced with the same quality permanently.
• Cowork is ideal as a "pre-stage to Claude Code." You can safely experience the power of AI agents.
• Work changes sufficiently with Chat alone. Using Projects/Memory/Deep Research makes it a different beast.
• No need to start everything at once. Start from the interface that suits the task you feel most stressed with.
• If you're only using one of the three, use this article as an opportunity to try the "neighboring face." That alone will change the quality of your work.

For those who found this article even slightly helpful:

The University of Tokyo Claude Code Laboratory (@ClaudeCode_UT) is an account run by a team of three University of Tokyo students. Joint development of AI operations with major companies is also underway.

We deliver "genuinely useful" information and know-how specialized for practical work every day.

■ Free distribution of Claude Code skills seriously developed by UTokyo students

■ Explanation of cutting-edge information from a field perspective

■ Provision of judgment criteria and Skills templates that can be used immediately in work

This is the only place where you can not only learn but also receive tools and skills seriously developed by UTokyo students for free.

Please also register for our LINE.

lin.ee/quYxNMc

If you are interested, please follow and check it out! ❗️