TSH Core

tsh-core holds what is neither a discipline's practice nor a stack's convention: Git worktree lifecycle, the house standard for writing technical documents, authoring Claude Code extensions, the project-context files Claude Code loads, the decision-record format and lifecycle, and a one-command project setup that wires them together. It also bundles three MCP servers: Atlassian for Jira and Confluence, Figma for design context, and Context7 for library documentation.

It is the only plugin in the marketplace defined by a negation (neither a discipline nor a stack), so it is deliberately small and has the highest admission bar in the repository.

Who installs it and where

Install it at user scope, like a discipline plugin: nothing in it depends on your role or on the language a repository is written in, so it should follow you everywhere. Stack plugins go at project scope instead, because the repository decides those. See Installation for the commands.

tsh-core is installed by people who have no other plugin at all, so its skills reference only files inside the plugin itself.

What it ships

Skill	Invoke	What it does
authoring-claude-extensions	/tsh-core:authoring-claude-extensions	Decides which Claude Code extension a need calls for — skill, subagent, hook, or plugin — and then writes it. Use when building or changing any of those four, when a workflow keeps getting pasted into chat, when a side task floods the conversation, or when a setup has to be shared across repositories or published as a plugin. A single hook wired into this repository's settings.json is Claude Code's own configuration, not an extension to author.
init	/tsh-core:init (on request only)	One-shot setup of a project's Claude Code context. Audits what already loads, then creates or repairs root and nested `CLAUDE.md`, path-scoped rules, and the decision-record archive with its index by running the owning tsh-core skills in order, and wires maintenance pointers into `CLAUDE.md` so future sessions keep it all current. Run /tsh-core:init inside the repository to set up.
managing-claude-context	/tsh-core:managing-claude-context	Sets up and maintains the project-context files Claude Code loads — root and nested `CLAUDE.md`, path-scoped rules in `.claude/rules/`, and a decision-record index — so a repository's conventions reach the model without bloating every session. Use when a repository has no `CLAUDE.md`, when Claude keeps ignoring a convention, or when the memory files have drifted from the code.
managing-decision-records	/tsh-core:managing-decision-records	Owns the shape and lifecycle of architecture decision records — the four-section format, the status vocabulary, numbering, superseding, and keeping `docs/decisions/README.md` in step. Use when adding a decision record, changing one's status, superseding an old decision, or deciding whether something warrants a record at all.
managing-git-worktrees	/tsh-core:managing-git-worktrees	Manages Git worktree lifecycle requests by creating a worktree from a freshly fetched base branch on origin — any branch, defaulting to origin's own default branch — listing registered worktrees, and removing a precisely identified worktree. Use when the user asks for worktree creation, listing, removal, or safe Git worktree lifecycle guidance.
writing-technical-documents	/tsh-core:writing-technical-documents	Applies TSH's writing standard to any technical document — README, CHANGELOG, ADR, RFC, runbook, PR description, ticket, user story, bug report or test plan — so it leads with the conclusion, states only verified facts, and is short enough to be read. Use when writing or revising a document meant for other people, or when asked to tighten, shorten or clarify prose.

The plugin ships no agents and no hooks. Every skill except init is model-invocable: Claude loads it when the work matches its description. init is the one deliberate exception (disable-model-invocation: true): it is a command you type, so its description is never preloaded.

How the skills divide the work

• Build something versus fix the memory files. authoring-claude-extensions owns skills, subagents, hooks and plugins; managing-claude-context owns CLAUDE.md, .claude/rules/ and the decision index. Each hands off to the other rather than overlapping.
• Setup versus upkeep. init is the command you type once per repository. It works by running managing-claude-context and its neighbours in order and wiring maintenance pointers into CLAUDE.md; every rule about the artifacts stays with the owning skill.
• Decision records are split three ways. The index schema, the binding note and the CLAUDE.md pointer belong to managing-claude-context; record format, numbering, status lifecycle and superseding belong to managing-decision-records; the prose inside a record belongs to writing-technical-documents. None of them sets a record's status: marking something Accepted stays a human call.
• Prose craft, never structure. writing-technical-documents governs the words inside a document, not which sections an artifact needs.

Bundled MCP servers

The plugin ships an .mcp.json with three remote HTTP servers. They appear in /mcp as plugin-provided and can be disabled per project from the same panel. Adding a server to tsh-core needs three disciplines that use it in a normal month; an MCP server adds nothing to the skill listing, and Claude Code defers tool schemas until a tool is used.

Server	Endpoint	Sign-in	What it gives Claude
atlassian	https://mcp.atlassian.com/v1/mcp/authv2	Run /mcp, pick atlassian, finish the browser login (OAuth)	Read and update Jira work items and Confluence pages directly. Atlassian Cloud only
figma	https://mcp.figma.com/mcp	Run /mcp, pick figma, finish the browser login (OAuth)	Read a design directly: layout, variables, components and a rendered export
context7	https://mcp.context7.com/mcp	None	Current, version-specific library and framework documentation, read-only

• Per-user access. Atlassian and Figma are OAuth-authenticated, so every call runs as you and grants nothing you could not already open yourself. Without the Figma login the server shows as needing authentication, and agents that need a design report what is missing rather than guessing at one.
• Context7 rate limit. Without an API key it works under Context7's anonymous rate limit; a teammate who hits it can add their own key in a user-scope context7 entry, which outranks the bundled one.
• Bitbucket and the Figma desktop server are not bundled. The same Atlassian server exposes Bitbucket Cloud only under API-token authentication, and the Figma desktop server is a separate local endpoint that needs the desktop app. The plugin README shows how to add each yourself.
• Do not pin tool names. A teammate's own user-scope entry wins on precedence and changes the tool namespace, so tool names are discovered at runtime and never written into an agent's tools:, a skill's allowed-tools: or a hook matcher.

For sign-in steps and server details see Connecting tools, Atlassian, Figma and Context7.

Why these servers live here

Each server is here because several disciplines use it and none of that changes with the language a repository is written in. According to the plugin README:

• Atlassian: engineering, product management and testing all read from Jira.
• Figma: product engineering judges implemented UI against it, and product management reads flows and annotations out of it during discovery.
• Context7: testing looks up the Playwright API, platform engineering reads Terraform provider, Kubernetes and CI documentation, and product engineering checks framework APIs while implementing.

What belongs in tsh-core

A contribution reaches tsh-core only by elimination: it would not change if the repository switched language or framework, and it would not change if the reader switched job. Core skills are tool mechanics, where the procedure is dictated by the tool's own semantics (git, the shell, Claude Code's memory subsystem), not by a TSH opinion. The README names two closed exceptions that wrap no tool, writing-technical-documents and managing-decision-records, and a pull request must name three of the five disciplines that would use an addition in a normal month and report the plugin's routing footprint (the combined name, description and when_to_use characters across its skills) before and after. A bundled MCP server adds no routing footprint, so for a server the three-discipline test decides. Ties go to a discipline plugin.

Changelog

Release history is in the tsh-core changelog.