claude code

this is claude code

a recording of Claude Code in the terminal.

Claude Code is Anthropic’s coding agent. it can inspect a repository, change files, run commands, and use the results to decide what to do next. you can work through a terminal, supported editor, desktop app, or web session.

ironically, Anthropic’s Claude Code feels the least personified to me out of Claude Code, Codex, and Grok Bot. the agents i’ve seen tend to have names describing their jobs: reviewer, infra engineer, things like that. Claude Code supports named agents; the role focused presentation is what stands out in my own use.

what happens after a prompt?

Anthropic describes an agent loop that gathers context, acts, and checks the result. the model chooses an action; the harness supplies tools, manages context, and returns what happened. a failed test becomes useful information for the next attempt.

for example, “fix the empty search result” might involve finding the search handler, reproducing the failure, changing the code, and running the relevant test. those steps can all follow from one prompt. the finish line still needs to be concrete: an empty result should render a useful message, and an ordinary search should keep working.

who’s running twenty agents?

sometimes you’ll see people doing this.

i did this for a TikTok and went through my five hour usage window in 15 minutes.

where does the work run?

terminal, editor, and desktop

the terminal and editor integrations keep Claude close to an existing development environment. selected code and editor diagnostics can help locate a problem; shell output and the final diff show what happened after it was located.

the desktop app adds a visual place for sessions, diffs, previews, and parallel work. a worktree gives a session its own checkout. two sessions that own different worktrees can still produce conflicting changes when their branches are combined, so file ownership and a final integration pass matter.

Claude Code in the desktop app. screenshot from Anthropic’s overview.

choose the interface that makes your next decision easy to inspect. a CSS change needs a rendered page; a parser change needs representative inputs and outputs. either can also need a code review.

web and Remote Control

web sessions, currently a research preview, run in a cloud environment. Remote Control lets a browser or phone steer Claude Code running on your machine. local tools, files, and configuration stay with that machine, and the local process needs to remain available for work to continue there.

picture a bug that only reproduces with your local development database. Remote Control can let you check on that investigation from your phone. a cloud task needs its own reproducible environment and access to the relevant data. choose the execution environment before choosing the screen you want to use.

what can i add or control?

Claude Code’s extension system gives different jobs to instructions, skills, tools, and events:

the task needs	the relevant piece
repository conventions and durable context	CLAUDE.md and scoped rules
a repeatable procedure	a skill
access to an external system	MCP tools and resources
a separate investigation	a subagent with a defined task
code to run at a lifecycle event	a hook
related extensions to travel together	a plugin

these pieces can compose. a repository instruction can point to a review skill; that skill can use an issue tracker connection and delegate an investigation. keep each piece small enough to explain what it contributes. adding a second place for the same rule gives you a second place to keep it current.

permissions control whether tool actions proceed. the Bash sandbox constrains filesystem and network access for commands inside it. the account behind an external tool adds another boundary: a service can only grant the access that identity has.

what should i get back?

it says it’s done. now what?

a useful session returns the change, the checks it ran, and whatever is still uncertain. “tests pass” needs the command and result behind it. a screenshot needs the route and state it shows. a proposed fix needs the behavior it is supposed to change.

for parallel work, add the checkout and owner. for work that moves between machines, add the environment and any state that stayed behind. these details make it possible to resume a task without guessing what the previous session meant by “done.”

Cowork

Cowork extends Claude into broader files and deliverables. an engineering task fits Claude Code when its output belongs in a repository and needs the project’s verification. a document or a collection of organized files has its own review criteria. name the output first, then pick the workflow around it.

the Cowork task interface in Claude. Anthropic’s Cowork guide.

the configuration guide shows where to put personal settings and shared project instructions. for a workflow you repeat, the extensions example builds a project skill and checks it against a real command.