choosing your setup

where should i start?

these are source based starting points for different workflows. the useful question is which arrangement makes your next task easy to steer and verify.

you want a visible task and diff workflow

begin hereClaude Code on desktop, with one repository and a check you can rerun

you already work through shell tools

begin herethe terminal, using the same scripts and branch you would inspect yourself

the task begins with selected code or diagnostics

begin herethe editor integration, with the failing behavior stated in the prompt

you need to steer a local task from another device

begin hereRemote Control, after checking host availability and account access

the task needs a separate cloud environment

begin hereClaude Code on the web, with reproducible setup and explicit network access

add one thing at a time

start with a repository instruction file, the project’s actual check commands, and permissions that fit the task. finish one useful change before adding a collection of skills, connections, and background agents.

then follow a repeated friction. if the same review procedure needs explaining every time, put it in a skill. if a task repeatedly needs a service, evaluate its MCP connection. if independent investigations keep filling the main conversation, try a bounded subagent. each addition should solve a problem you can name and produce an outcome you can inspect.

how do i compare two setups?

a small repeatable exercise makes comparisons useful. use the same repository revision, the same request, and the same acceptance criteria in each setup. record the surface, model, permissions, and extra context supplied to each run.

for example, ask two fresh sessions to fix the same input validation bug. check whether each found the actual failure, preserved the intended behavior, ran a meaningful test, and left a readable diff. note the interventions you made. a faster first answer may still leave more review work.

one task gives evidence about that task. try a different kind of work before generalizing from the result: investigation, implementation, and review can favor different arrangements.

would more agents help?

an interface change and a documentation update can progress together once the behavior is agreed. two workers redesigning the same module need more coordination. use separate worktrees for independent edits and assign one owner to combine and verify the result.

start with a small number of tasks you can actually review. track those waiting on decisions as carefully as those still running. useful parallelism ends in work you can integrate.

Boris Cherny shared his own setup in January 2026: five Claude sessions in numbered terminal tabs, with notifications when one needs input. the useful detail is how he keeps track of work that needs his attention.

Boris Cherny (@bcherny), january 2026.

should i rely on a preview?

Claude Code’s agent guidance includes experimental agent teams. web sessions and routines remain a research preview. the Function Hooks discussion points toward deeper customization. decide how much change a workflow can absorb before depending on one of those interfaces.

use a disposable project for an experiment, capture the version and expected result, and keep a way to return to the supported setup. a reproducible small test gives a feature somewhere concrete to prove useful.