Anthropic: Put the tool where the work already happens.
Anthropic released Claude Code as a limited research preview that works inside the developer’s terminal and keeps them in the loop. Explore why where a tool lives, and who approves its actions, can matter as much as what it can do.
See the idea in one picture ↓- research preview
- 24 Feb 2025
- generally available
- 22 May 2025
- the developer approves
- In the loop
Figures as reported by the company.
Independent educational analysis. Scenarios, people, numbers and illustrations on this page are original PMcademy practice material, not company research results or an endorsement.
From copy and paste to working in the project
- 1Copy, paste, repeat
Moving code between a chat and a project is slow and loses context.
- 2Work where the code is
In the terminal, the tool can read, edit, test and push in the project itself.
- 3Ask before acting
The developer stays in the loop, and the preview became generally available.
Animated illustration. Names, numbers and people are fictional teaching material.
A good answer still leaves the work to you.
Copy, paste, run, repeat
The tool works where the code lives
Read the analysis
Imagine the developers on your team already ask a chat assistant for code. Each answer has to be copied into the project, run, fixed and run again. The assistant never sees the files, the tests or the error message. The product question is not only how good the answers are. It is how many steps separate an answer from finished work, and who carries the context across each step.
Fewer hand-offs need more control.
Illustrative data and simplified behaviour. This model does not represent a live company system or measured company results.
Read the analysis
Use the model below to compare three set-ups. In the first, the person copies text between a chat window and their project. In the second, the assistant works inside the project and can read files and run tests. In the third, it also asks before risky actions. Notice that removing hand-offs also increases what the tool could break, so control has to grow with capability. The step counts are illustrative.
Meet people in the place they already work.
“I do not want to explain the whole project to another window.”
Read the analysis
Our fictional learner, Mei, is fixing a failing test late on a Friday. Her terminal is open and the project is in front of her. Switching to another app means re-explaining the problem and pasting results back by hand. Ask where your users spend their working day and what they would have to carry from one tool to another. That gap is often the real cost your product can remove.
Your own team is the first test, not the last.
- What Anthropic says
Anthropic launched a limited preview, kept developers in the loop and later made it generally available.
- What it cannot prove
How much time it saves for teams outside Anthropic.
- What you would measure
Tasks delegated and finished
Approvals declined
Changes undone afterwards
Read the analysis
Anthropic said the tool had become indispensable for its own team, especially for test-driven development, debugging and large refactors, and that it released the preview to learn how developers use it. Internal use exposes rough edges quickly and shows which tasks people come back for. Colleagues, however, share context, patience and skills that customers may not. Pair internal use with an outside preview, and watch which tasks people delegate, where they step in and what they undo.
Start narrow, then widen the surfaces.
- Fewer hand-offs
- Fits existing habits
- Works with other tools
- Harder for newcomers
- More it could break
- Permissions to design
Read the analysis
A terminal tool fits developers’ existing habits and scripts, and it can be combined with the other tools they use. It also asks more of a newcomer than a polished app full of buttons. When Claude Code became generally available in May 2025, Anthropic also announced beta extensions for VS Code and JetBrains and background tasks through GitHub Actions. Starting with one surface and adding others later is a sequencing choice. Decide which surface proves the value first.
Where should version one live?
Your fictional team is building an assistant for financial analysts. You can ship a standalone web app, a plugin for the spreadsheet they use all day, or a command-line tool for the most technical users. Choose where version one lives and what it may do without asking. Explain what you would learn first and what you are deliberately postponing.
Choose a next move
Leave with a hand-off map.
Your artifact
Write a hand-off map and a permission list: what the tool may do freely, what needs approval, and what it must never do.
Critique my project artifact using only the evidence I provide. Write a hand-off map and a permission list: what the tool may do freely, what needs approval, and what it must never do. Separate observed facts, assumptions, proposed tests and missing evidence. Identify a failure mode and a cheap check. Do not invent company results or user research. Ask for my artifact before starting.
Read the analysis
Map today’s workflow step by step and mark every hand-off between tools. Choose where your product should sit to remove the most costly one. List the actions it may take freely, the ones that need approval and the ones it must never take. Add how you will learn from your own team first, and what evidence would make you widen access.