claude-pro-kit

Runs search-only subagents (Explore by default) on Haiku instead of your main model; a model Claude asks for still wins, and /agent-diet picks the types and the model.

Explain, plan and ELI5 pages drawn natively in a side pane; plan decisions as buttons, Respond fills the prompt box.

Holds a prompt back once your 5-hour or weekly usage is at or over your limit (90% by default); sending it again goes through.

Status line countdown until the prompt cache expires; once it has, the exact tokens your next message re-sends uncached.

Adds quiet flags to noisy shell commands before they run (git status, pytest, cargo, npm install, mvn, curl, wget, docker pull), so Claude reads short output from the start; errors, failures and warnings stay in full.

Asks before Claude edits a file another chat on this machine changed in the last 30 minutes.

After a compaction, adds a note of exact facts from before it (your latest prompt in full, up to 4 earlier ones, the todo list, the last failed shell command, files edited), at most 4,000 characters, with no model call.

/xray opens a pane with the exact breakdown of what fills the context window and which tools load on every request.

Runs each git push, pull, fetch, clone and gh command with the token of the logged-in GitHub account that can see the repository's owner, without switching the active account.

Installs updates to your claude-pro-kit mods by itself at session start, and announces new mods (or installs them too, with /kit-update auto-new on); /kit-update installs updates and new mods now.

Holds back a shell command once when the exact same command already failed twice in a row, and tells Claude to change approach; retrying it still goes through.

/nudge <goal> applies behavioral design to a product or flow: diagnoses why users don't act, designs interventions that remove barriers, ranks hypotheses to test, and checks every idea for ethics. Its guide reaches Claude only when you run /nudge.

Trims long shell output, Grep results and subagent reports before Claude reads them (head, tail and error lines kept); the untrimmed text is saved to a file Claude can open.

Typing "status" while background shells or subagents run opens a pane with each task's exact elapsed time and last output lines instead of sending the prompt to Claude; /peek opens it too.

Desktop heads-up display: live session, week and context meters, per-turn token receipt, usage in the spinner, compact tool rows.

An Improve button above the prompt (and /polish) rewrites your draft with Opus at low effort (or Haiku or Sonnet) using the prompt-master rules and puts it back in the box, never sent; each press is one model call, with its exact tokens shown.

Stops a whole-file read of a file over 1,000 lines once, with Read, cat or Get-Content, and gives Claude an outline tool so it reads only the range it needs.

Skips Claude's identical re-read of a file that has not changed, with Read or a plain cat, sed -n, head, tail or Get-Content; a deliberate retry still goes through.

/receipt opens a pane with the exact tokens every turn of the session spent and the costliest turns.

Lists skills you have not used lately in this project by name only, without their descriptions; the Skill tool and /name still load any of them.

Loads tools you have not used lately on demand through ToolSearch instead of with every request; core tools and recently used ones stay loaded.

When one turn crosses +5 session points, writes a handoff with your latest prompt in full and copies it for a new session; the old session is never cleared. /handoff any time.

Steers Claude to Edit instead of rewriting an existing file in full: the first rewrite in a conversation gets a one-line note, a later one is held back once; a deliberate retry still goes through.