Creator / AI Workflow Builder· 2026

Claude Code Toolkit: A Mod Collection for Tidier Long Sessions

TypeScriptProductivityClaude CodeOpen SourceClaude Code Mods

Overview

A pegboard of small tools, with a Fat Invader in a tool belt hanging on a hook among them

These mods came out of my own long days in Claude Code: several sessions open on the same project, a context window that fills up, answers I wanted shorter or plainer. Each one fixes one small thing that kept going wrong, and together they are the Claude Code Toolkit.

The mods

context-alarm warns when the context window passes 80%, again at 90%, and suggests saving the session's decisions with /save-context before automatic compaction summarises them away.

dash-guard refuses prose Claude writes with an em dash, an en dash or a double hyphen used as punctuation, and tells Claude which lines to rewrite.

dash-guard refusing a Write to notes.md that holds an em dash (an example session, rendered from the mod's own output) — dash-guard: the write is refused, with the line to rewrite and how.

shared-file-guard is for running several sessions on the same project. It refuses a shell command that would write a shared file such as STATUS.md, CLAUDE.md or MEMORY.md when this session never read it, or another session changed it since.

shared-file-guard refusing a shell write to a status file that another session changed (an example session, rendered from the mod's own output) — shared-file-guard: another session changed the file, so the heredoc write is held until the file is read again.

answer-buttons draws small buttons under Claude's last answer: Plain English, Shorter, and Plain and short, so a long answer becomes the one I needed in one click.

skill-stats shows how often each of your skills runs, which ones never trigger and which reach Claude without their description, and hands a weak description to skill-creator to fix.

skill-stats listing use counts, never triggered skills and a skill listed by name only (an example session, rendered from the mod's own output) — skill-stats: use counts from history and live, the skills that never trigger, and one listed by name only.

subtask-icons puts one small Subtask button under Claude's last answer. It opens a picker of the answer's items, and a pick puts /subtask with that item in the prompt box, so one point of a long answer can become its own piece of work.

Beside it, a Pin list button pins that answer's items in a column above the prompt, so the list stays in view however far the conversation scrolls. Nothing pins by itself. Items I send with /subtask get a ⑂ mark, /st pins shows the last few answers that had a list so an earlier one can be pinned with /st pin N, and /st unpin clears it.

fork-lineage came from my own forks. I would fork a session to try something on the side, rename the fork, rename the parent, and soon nobody knew which came from which, and whatever the fork had worked out stayed in the fork. Now the band above the prompt says which session a fork came from, by its current name, so renames never break the link, and /lineage shows the whole family.

An animation of fork-lineage: a session forks, both are renamed and the link holds, the fork works, a report is approved and flies back to the parent's mailbox, and the parent reads it when it comes back — fork-lineage: the link survives renames, and the report goes back only after you approve it.

When the fork's work is done and I want to close it, a report to parent button asks the fork's own model for a short report of what was decided, changed and left open. I see it and approve it before anything goes. If the parent session is open, it gets the report at once. If it is closed, the report waits, and when I resume the parent a blinking notice and a Read button bring it into the conversation. Nothing is ever sent on its own.

A recording style animation of fork-lineage from /branch to Read: the fork works over several turns, reports after approval, an open parent receives it at once, a closed one finds it waiting (an example session, rendered from the mod's own output) — fork-lineage from start to end: fork with /branch, work, report before closing, and the parent receives it, now or when it comes back.

Install

One command installs every mod in the Toolkit, or install only the ones you want. Nothing is tracked and nothing leaves your machine. You need Claude Code 2.1.287 or later; fork-lineage also needs 2.1.289 and python3.

bash

/plugin marketplace add mutlumehmet/claude-plugins
/plugin install toolkit@mehmetmutlu

# or one mod
/plugin install fork-lineage@mehmetmutlu

Open source on GitHubmutlumehmet/claude-plugins/plugins/toolkitView the code