What's 'Worth Doing' and 'What Can Be Skipped' with Claude Code: Categorizing 12 PM Tasks into 3 Tiers
In 5 minutes, you will have a classification table that changes your Claude Code launch frequency from 'once a week' to 'five times a week' starting tomorrow.
You thought, 'I can do everything with Claude Code,' but ended up rarely opening it.
For those who felt that way: last time (Article 011), I organized the reasons for not opening it into three categories: 'long startup ritual / justification gap / lack of tasks.' This is the follow-up.
To give you the conclusion first, out of 12 PM tasks, '8 are suited for Claude Code, 3 are better for the browser (Cowork), and 1 is fine with ChatGPT,' meaning there are more situations where Claude Code is useful than you might imagine. It wasn't that 'there are no situations to use it,' but rather that 'you just haven't noticed the situations to use it.'
Claude Code is an agentic AI coding tool provided by Anthropic that runs in the terminal (CLI). It can cross-process not only codebases but also local Markdown files, meeting minutes, PRDs, long interview transcripts, and **'anything you have on hand as a file'** (as of April 2026).
In this article, I will categorize the 12 tasks that a BtoB SaaS PM performs daily into three tiers: 'Suited for Claude Code / Browser-oriented / Fine with ChatGPT.' By the time you finish reading, you should know the first command to run tomorrow morning.
1. 'Everything in Claude Code' doesn't last
Let's briefly review the three frictions from last time once more.
| # | Friction | Essence |
|---|---|---|
| 1 | Long startup ritual | The 4-step process of iTerm launch → cd → claude doesn't match the speed of thought |
| 2 | Justification gap | I don't have an explanation for myself on 'why a PM uses CLI' |
| 3 | Not opening due to lack of tasks | Mismatch with an idea-first work style |
If you try to solve these by 'opening it with willpower,' you won't open it again next week.
The premise for a solution is to pass only tasks that justify the startup cost to the CLI. If you try to bring even spur-of-the-moment questions to Claude Code, you will be exhausted by the 4-step ritual every time. Conversely, you only need to run tasks that are 'worth paying the startup cost for,' such as 'cross-referencing many files,' 'assigning long-duration processing,' or 'structuring local documents,'tasks worth paying the startup cost for.
To do that, we start by dividing your work into 3 tiers.
2. What is the 3-tier classification model?
The 3-tier model is a framework for dividing what to entrust to which AI based on startup cost and processing characteristics.
2-1. Characteristics suited for Claude Code
Looking across / cross-editing many files
Processing Markdown / PDF / code / logs located locally
Running long-duration (several minutes or more) processing
Wanting to save the output as a file
Specifics: Organizing PRD templates, structuring past meeting minutes, creating local comparison tables for competitor materials, and drafting diff comments for PR reviews.
2-2. Browser-oriented (e.g., Cowork) features
Summarizing browser articles and videos
Drafting emails and chats
Consulting while screen sharing
Tasks that involve repeated copy-pasting via the clipboard
Specifics: Summarizing Slack discussions, live viewing and comparing competitor sites, real-time verbatim transcription of customer interviews, and creating drafts to post to internal Wikis.
2-3. Areas where ChatGPT is sufficient
Spontaneous, one-off questions
Using as a dictionary
Rewriting text that ends in a single exchange
One-off translations
Specifics: Questions that can be finished in 5 minutes, such as 'How to use VLOOKUP,' 'Phrasing for this English sentence,' or 'Polishing OKR wording'.
The purpose of dividing into 3 tiers is to stop trying to 'do everything' and instead choose a way of working that matches the startup cost.
3. Actually distributing 12 PM tasks
I will list 12 tasks I perform daily as a BtoB SaaS PM and distribute each into the 3 tiers. The same distribution should work for your tasks as well.
3-1. Organizing meeting minutes (50 Slack messages → 3 lines of decisions) → Best for Claude Code
Paste a long Slack thread into a local `.txt` file and ask Claude Code to organize it into 3 tiers: 'Decisions/Actions/Unresolved.' You can compress 50 messages down to 3 lines. Since it's highly valuable to turn into a file, CLI is the only choice.
3-2. Requirement definition draft (1-page PRD) → Best for Claude Code
Pass a PRD template MD to Claude Code and have it fill in 'Functional Requirements, Non-functional Requirements, Out of Scope, and Assumed Risks.' The ability to cross-reference your company's templates and past PRDs is Claude Code's strength.
3-3. Competitor research (Comparison table of 5 companies x 3 axes) → Browser-oriented (Best for Cowork)
Since I compare the official pages of 5 companies while keeping them open in the browser, Cowork, which is integrated with browser operations, is the best choice. While Claude Code can also fetch HTML from official pages, Cowork is stronger when it comes to checking UI changes and interactive elements.
3-4. Refining OKR wording → ChatGPT is sufficient
For tasks that can be completed in a single round-trip, such as 'Rewrite this OKR to be a bit more quantitative,' ChatGPT is sufficient. There is no value in launching Claude Code.
3-5. Transcribing customer interviews → Best for Claude Code
Place audio files (or transcribed MD files) locally and execute 'organize by segment -> extract key phrases -> rank by importance' in one go. Even if processing takes several minutes, the CLI is the best because it completes with file output.
3-6. Converting roadmap to Markdown → Best for Claude Code
Manage quarterly roadmaps locally, such as in `roadmap-2026-q2.md`. If you ask Claude Code to 'show the difference between Q1 and Q2' or 'list only the items not yet started,' file management becomes much easier.
3-7. Summarizing Slack discussions → Better for browser
While keeping Slack open in the browser in real-time, ask Cowork to 'summarize the last hour of discussion into 3 lines.' Since you use it while the discussion is ongoing in the browser, Cowork is the choice.
3-8. Primary research acquisition (fetching official documentation) → Better for browser
'Fetch the latest XX official documentation and extract the changes' is better suited for Cowork, which includes browser navigation. Claude Code alone would just hit the URL directly, and it is weak at deep-dive searches that require navigation.
3-9. Creating diagrams (mermaid / ASCII) → Best for Claude Code
Creating diagrams using mermaid notation in files like `flow.md` is a job for Claude Code only. Diagrams that can be coded are most efficiently written via CLI.
3-10. PR review (diff comment suggestions) → Best for Claude Code
Launch `claude` within the Git repository and have it provide comment suggestions for the `git diff` content. Git integration and cross-file analysis are the areas where Claude Code is strongest.
3-11. Maintaining internal FAQ → Best for Claude Code
Place MD files exported from Notion into an `faq/` folder and ask Claude Code to 'remove duplicates,' 'categorize,' and 'suggest missing categories.' Organizing across multiple files is a job for Code only.
3-12. Error investigation (checking bugs in our own product) → Best for Claude Code
Save log files locally and ask Claude Code to 'classify error patterns' and 'extract reproduction conditions.' File input and long-duration processing are typical examples where Code shines.
4. Summary: Out of 12 tasks, 8 are CLI-oriented, 3 are browser-oriented, and 1 is sufficient with ChatGPT
I will summarize the sorting results in a table.
| # | Task | Optimal Tool |
|---|---|---|
| 3-1 | Meeting Minutes Organization | Claude Code |
| 3-2 | Requirements Definition Draft (PRD) | Claude Code |
| 3-3 | Competitive Research (5 companies x 3 axes) | Cowork |
| 3-4 | OKR Wording Refinement | ChatGPT |
| 3-5 | Customer Interview Transcription | Claude Code |
| 3-6 | Roadmap Markdown Conversion | Claude Code |
| 3-7 | Slack Discussion Summary | Cowork |
| 3-8 | Primary Research Acquisition | Cowork |
| 3-9 | Diagram Creation (mermaid) | Claude Code |
| 3-10 | PR Review (diff) | Claude Code |
| 3-11 | Internal FAQ Maintenance | Claude Code |
| 3-12 | Error Investigation | Claude Code |
Total: Best for Claude Code 8 tasks / Best for Cowork 3 tasks / Sufficient with ChatGPT 1 task
In other words, two-thirds of a PM's work can be handed over to Claude Code. The idea that "there's no situation to use it" was a structural misunderstanding.
Estimated launch frequency: 3-5 times a week (operating by batching tasks). You can significantly increase this from once a week.
5. The first command to run tomorrow morning
Once you see the sorting, decide on the first command to run tomorrow morning. Just "opening it for now" won't last, so decide on just one command where the initial launch becomes a "means" rather than an "end".
I recommend **"summarizing unread Slack messages"**.
# 朝の最初の1コマンド例
cd ~/work/notes
claude "今朝のSlack未読をまとめて、決定事項とアクションだけ抽出して"Once this morning command becomes a habit, launching changes from an "end" to a "means." It becomes "I need to organize Slack → launch Claude Code" instead of "I have to open Claude Code." The psychological friction of launching no longer comes before the goal.
6. What about the 3 browser-oriented tasks?
Many of you probably think, "I don't use Cowork." Actually, without Cowork, those 3 browser-oriented tasks tend to flow to ChatGPT, which creates the illusion that **"ChatGPT is all-powerful after all."**
However, because ChatGPT is not integrated with browser operations, the number of copy-paste round trips increases, which ends up wasting time.
Next time (Article 013, scheduled for release on 5/2), I will explain the procedure for introducing Claude Cowork as a partner to Code and assigning the 3 browser-oriented tasks to Cowork. This completes the division of roles between Code (CLI) and Cowork (GUI).
7. Frequently Asked Questions
Q1. Can I use the 12-task sorting for other professions, not just PMs?
Yes, it is a framework that can be used regardless of profession. The content of the tasks differs by profession, but anyone can apply it by sorting them along the three axes of "cross-referencing many files," "involving browser operations," and "finishing in one round trip." Whether you are an engineer, marketer, customer success representative, or in legal, try writing down your own 12 main tasks.
Q2. Out of the 12 tasks, which one should I do first?
I recommend "Meeting Minutes Organization (3-1)" or "Slack Discussion Summary (3-7)." Both have high frequency and you can feel the effect immediately, which will rapidly increase your trust in Claude Code. The experience of "I can use this too" will drastically lower the psychological friction of launching.
Q3. Was it a lie that Claude Code can "do everything"?
It is more accurate to say that "being able to do it" and "being suited for it" are different. Claude Code can technically handle all tasks, but the reality is that it also includes work that isn't worth the launch cost. The 3-tier sorting is a concept of "choosing situations where Claude Code can be utilized" and does not deny the capabilities of Claude Code.
Q4. Does this sorting still hold up if I don't have a ChatGPT subscription?
Yes, it does. The one task intended for ChatGPT (polishing OKR wording) can also be handled via the Claude.ai web chat. Alternatively, you can do it in Claude Code without any issues (though the value of paying the startup cost is slightly lower, it's sufficient if you don't have ChatGPT). The important thing is how you handle the '3 browser-oriented tasks'.
Q5. Isn't this sorting too rigid? Won't it change depending on the situation?
Yes, it will change depending on the situation. The sorting in this article is 'my workflow as of April 2026,' and there is a strong possibility that the 3 browser-oriented tasks will be integrated into Code through Cowork updates or new Claude Code features. The important thing is not the 'specific answer' of the sorting, but the framework of 'thinking in 3 layers'. Once you grasp the framework, you can re-sort them yourself even as tools evolve.
8. Summary: Two-thirds can be handed over to Claude Code
Let's briefly summarize what we've covered so far.
When you sort the 12 PM tasks into 3 layers—'Claude Code-oriented / Browser-oriented / Sufficient with ChatGPT'—you'll find that 8 tasks (67%) are suited for Claude Code
It wasn't that there were 'no situations to use it,' but rather that I hadn't noticed the situations to use it.
By deciding on the first command to run tomorrow morning in advance, the psychological friction of starting it up is reduced.
The 3 browser-oriented tasks will be assigned to Cowork in the next installment (Article 013).
By the time you finish reading the 3 articles, your Claude Code startup frequency should have changed from once a week to five times a week. Next time, we will cover the procedure for handling the 3 browser-oriented tasks with Cowork.
If you'd like to read next
I set up Claude Code but only open it once a week—The blind spot called 'psychological friction of terminal startup' — Series #1 (Previous article: Articulating the pain)
Claude Code's partner was Claude Cowork—The second AI that 'allows' browser-oriented work (Article 013, scheduled for release on 5/2)
[2026 Latest] What is Claude Code? 5 ways to triple your productivity — An introductory article on Claude Code itself
This article is based on operations as of April 2026. Claude Code / Claude Cowork features may change in the future. Please check the official Anthropic documentation for the latest information.
If you're wondering, 'How did my 12-task sorting turn out?', please let me know in the comments section or on Threads. I'd like to use it as a case study next time.