Technology anatomy
What is this built on?
Pick a tool, then dive downward. The magical surface is usually sitting on a few ordinary layers: protocols, APIs, files, shells, operating systems, and human decisions.
Current stack
Claude Code
How Claude Code turns a developer request into file edits, terminal commands, and verified code changes.
Claude Code gives the developer a conversational interface inside the terminal. The visible experience feels like asking an expert pair programmer, but the work happens through concrete operations: reading files, searching code, editing text, running commands, and checking results.
Built on
- Project context
- Developer instructions
- Terminal session
Enables
- Codebase exploration
- Refactoring help
- Feature implementation
Examples
- Ask Claude to fix a bug
- Review a diff
- Run tests after edits
The agent loop is what makes Claude Code feel active. Claude reasons over the request and repository state, selects a tool call, receives structured output, then repeats until it can explain, edit, test, or stop.
Built on
- Language model
- Tool descriptions
- Conversation history
Enables
- Multi-step tasks
- Error recovery
- Iterative debugging
Examples
- Search → read → edit → test
- Build → inspect error → patch
Alongside the local tools is a large language model running inference. It turns tokens into probable continuations: explanations, plans, code edits, or tool requests. The surrounding agent system gives those predictions structure and feedback.
Built on
- Transformer model
- Context window
- Inference servers
Enables
- Reasoning over code
- Instruction following
- Tool-use planning
Examples
- Claude Sonnet
- Context tokens
- Generated patches
Claude does not directly “touch” your computer. It asks for specific tool calls such as reading a file, applying an edit, listing search results, or running a shell command. Each tool result becomes new context for the next decision.
Built on
- Tool schemas
- Runtime permissions
- Serialized outputs
Enables
- Filesystem access
- Shell access
- Auditable edits
Examples
- Read files
- Search with rg
- Apply a patch
Most of Claude Code’s practical value comes from coordinating standard development tools quickly. It can inspect the repository, understand conventions, run tests, read stack traces, and use git status or diffs to keep changes visible.
Built on
- Source files
- Package manager
- Test runner
- Git repository
Enables
- Verification
- Project-aware edits
- Change review
Examples
- npm test
- git diff
- TypeScript errors
At the bottom is the same operating system every developer tool uses. It manages files, launches processes, handles environment variables, enforces permissions, and connects software to hardware and networks.
Built on
- Kernel
- Process model
- Storage
- Permissions
Enables
- Terminals
- Editors
- Package managers
- Development servers
Examples
- Processes
- Environment variables
- File permissions
Claude Code only becomes useful when it reaches a person through sensory and action channels. You see streamed answers, file paths, diffs, test results, and errors. You type requests, approve directions, notice surprises, and decide what deserves another check.
Built on
- Screen or terminal display
- Keyboard or voice input
- Attention and review
Enables
- Reading generated changes
- Steering the agent
- Catching mistakes before shipping
Examples
- Scan a diff
- Type a follow-up prompt
- Notice a failing test
The deepest layer is not another server. It is the person with a job to be done. For Claude Code, that person brings the project goal, judgment, taste, constraints, and responsibility for deciding whether the result is good enough.
Built on
- A project goal
- Developer judgment
- Real-world constraints
Enables
- Meaningful prompts
- Useful acceptance criteria
- Responsible decisions
Examples
- Ship a feature
- Fix a bug
- Learn how a codebase works
