ira-claude-plugins

Personal Claude Code skills, published as an installable plugin marketplace.

Description follows below after the installation instructions.

Install in ClaudeCode

/plugin marketplace add seefood/ira-claude-plugins
/plugin install test-before-install@ira-claude-plugins
/plugin install json-to-toon@ira-claude-plugins
/plugin install debugging-with-the-scientific-method@ira-claude-plugins
/plugin install settings-curator@ira-claude-plugins

Install with homeshick in Kiro and Cursor

This repository is also a homeshick castle. With homeshick installed, clone and link it like this:

homeshick clone git@github.com:seefood/ira-claude-plugins.git
homeshick link ira-claude-plugins

Linking the castle installs the global skills and Kiro permission hook from home/:

• ~/.cursor/skills/ — Cursor copies of the skills
• ~/.kiro/skills/ — Kiro copies of the skills
• ~/.kiro/hooks/permission-gate.json and ~/.kiro/hooks/permission-gate.py — a Kiro PreToolUse hook that applies Claude-style allow/deny rules from the current project's .claude/settings.json

The hook falls through to Kiro's built-in permission system when no matching Claude rule exists. Linking the castle manages all files under home/, not just these paths; use homeshick unlink ira-claude-plugins to remove its links without deleting the castle checkout.

Install as a Kiro Power

This repository is also a Kiro Power using the vendor-neutral Agent Plugins format. The root plugin.json identifies the Power, and the root skills/ directory exposes the existing skills under plugins/ without duplicating their files:

plugin.json
skills/<name> -> ../plugins/<name>/skills/<name>

Each symlink points to the complete skill directory, including its SKILL.md, references/, and scripts/ files. The Agent Plugins manifest has no path-remapping field, so these root-level links provide the conventional Kiro skills/<name>/ layout while keeping the Claude plugin sources canonical.

Install as a Kiro Power

In Kiro, open the Powers panel, choose Add Custom Power → Import power from GitHub, enter:

https://github.com/seefood/ira-claude-plugins

and install it. Kiro can update a GitHub-installed Power from its source; see Install Powers and Create Powers.

The Kiro Power packages the four skills. The Claude-compatible permission gate and other global files under home/.kiro/ remain managed by the homeshick installation above, because Kiro Powers do not currently document packaging arbitrary global hooks or configuration.

Plugins

test-before-install

Vet a third-party agent artifact (Claude Code skill, plugin, MCP server) or a local "enhancement" from a remote or external source before installing it locally. Fetches the source read-only into quarantine, runs static and type-appropriate scanners in isolation, verifies findings by hand, and reports a verdict with the side effects to consent to.

See plugins/test-before-install/skills/test-before-install/SKILL.md.

json-to-toon

Converts JSON command output (AWS CLI, kubectl -o json, terraform show -json, snyk-agent-scan --json, etc.) to compact TOON notation before it enters the conversation, to save ingestion tokens. Skips conversion when the raw JSON itself is the deliverable (editing, piping to another JSON consumer, saving for later). Maintains a living registry of known JSON-emitting commands that grows as new ones are discovered.

See plugins/json-to-toon/skills/json-to-toon/SKILL.md.

Measured token savings on real command output (tiktoken, cl100k_base / o200k_base — an approximation of the actual Sonnet 5 tokenizer, not an exact count):

Savings scale with how tabular the data is: uniform arrays of objects collapse into key[N]{fields}: header + rows, while prose-heavy fields (e.g. long content strings) don't compress much beyond removing JSON punctuation.

settings-curator

Consolidate scattered per-directory .claude/settings.local.json files (created when Claude Code sessions were started from subdirectories instead of the repo root) into the root .claude/settings.json as durable, curated permission rules, then replace the redundant nested .claude directories with symlinks to the root one so future sessions share one config.

See plugins/settings-curator/skills/settings-curator/SKILL.md.

debugging-with-the-scientific-method

Replaces ad hoc "try this, try that" debugging with a written lab-notebook log of hypothesis, experiment, and result — one falsifiable hypothesis at a time, forcing precision and preventing repeated dead ends.

See plugins/debugging-with-the-scientific-method/skills/debugging-with-the-scientific-method/SKILL.md.

Method and example adapted from Noam Elfanbaum's debugging-with-the-scientific-method.

Repository layout

The repository combines three packaging formats:

Claude Code marketplace:

.claude-plugin/marketplace.json                 - marketplace manifest
plugins/<name>/.claude-plugin/plugin.json       - per-plugin manifest
plugins/<name>/skills/<name>/SKILL.md           - canonical skill files

Kiro Agent Plugin / Power:

plugin.json                                     - Agent Plugins manifest
skills/<name> -> ../plugins/<name>/skills/<name> - skill directory links

The root skills/ links expose the complete skill directories, including SKILL.md, references/, and scripts/, without duplicating the canonical files under plugins/.

Homeshick castle:

home/.cursor/                                   - Cursor global files
home/.kiro/                                     - Kiro global skills, hooks, and settings

See the Claude Code plugin marketplace schema, the Agent Plugins specification, and the homeshick documentation for the corresponding packaging formats.

License

MIT, see LICENSE.