Practical Work Kit

Four AI skills: check work, challenge plans, shape ideas, and carry context.

Four reusable AI skills for Codex and Claude: check work, challenge plans, shape ideas, and carry context into the next conversation. Both hosts use the same skill files.

Examples · Latest release · License

Give the assistant the material you already have and say what you need. The skills aim to produce useful answers at the requested scale: a short message gets a short review, three titles stay three titles, and a handoff preserves the draft you actually wrote.

Skill	Use it when you need…	Example request
`check-my-work`	Errors, omissions, and evidence behind a finished result	“Check this budget and tell me what needs fixing.”
`challenge-my-plan`	A second perspective on a plan and its assumptions	“Challenge this launch plan and suggest a smaller first test.”
`shape-my-idea`	A concrete direction, draft, or first experiment	“Turn this idea into three video concepts I can make this weekend.”
`carry-my-context`	A concise note for another person or conversation	“Keep my decisions, current draft, and next step in a handoff.”

See it in one minute

Try a small budget check after installation. Examples use Codex's $skill-name; in Claude, select the matching skill with /practical-work-kit:skill-name, then paste the same request.

Use $check-my-work to check this budget.
Venue: 3,000. Materials: 2,000. Listed total: 6,000.
I have not provided receipts or confirmed tax.
Give me a short correction and say what remains unchecked.

What a useful answer should do: correct the total to 5,000, flag the 1,000 difference, and leave receipts and tax marked as unverified. You can immediately fix the arithmetic and see what still needs checking.

This is a worked example, not a guaranteed model response. See four complete input/output examples to try the other skills.

Install in Codex

Use a Codex CLI with the plugin commands available:

codex plugin marketplace add Yingye107/practical-work-kit --ref main
codex plugin add practical-work-kit@practical-work-kit

Start a new conversation after installation. To inspect the installation:

codex plugin list --marketplace practical-work-kit --json

For a fixed release, use --ref v0.2.1 instead of --ref main. If your CLI does not recognize plugin, see the official marketplace documentation for your host and version.

Install in Claude Code

Run these commands in your terminal:

claude plugin marketplace add Yingye107/practical-work-kit
claude plugin install practical-work-kit@practical-work-kit

Start a new session, then use /practical-work-kit:check-my-work, /practical-work-kit:challenge-my-plan, /practical-work-kit:shape-my-idea, or /practical-work-kit:carry-my-context. To check the installation, run claude plugin list. See Claude Code's official installation guide.

Install in Claude chat or Cowork

In Claude's Customize → Plugins, choose Add → Add marketplace and enter Yingye107/practical-work-kit. Alternatively, choose Add → Upload plugin and upload the release ZIP. In chat, type / and select a skill from Practical Work Kit. These options depend on your account and organization settings; see the official Claude guide.

An installation from the Claude Code terminal stays on that machine; it does not add the plugin to your Claude.ai account.

Download or use a fixed version

Get practical-work-kit-0.2.1.zip and SHA256SUMS.txt from Releases. The same ZIP includes both host manifests and all four skills. For a fixed Claude Code version, extract it into a dedicated folder and use the local marketplace instructions in INSTALL.txt.

GitHub publishing does not mean approval or listing in OpenAI's or Anthropic's public directories. Claude Code installation and skill discovery were checked locally; model responses and Claude's account-based installation have not been tested by this release.

Install one source of these skills. If you already use a local copy or individual versions, avoid enabling duplicate copies alongside this bundle.

Try it

Paste your material after an explicit skill request:

Use $check-my-work to check this budget.
Venue: 3,000. Materials: 2,000. Listed total: 6,000.
Tell me whether the calculation is right and what was not checked.

Use $shape-my-idea to give me three titles for a five-minute video
about organizing a small work desk. Keep the tone friendly.

You can ask in your own language and set the length. Explicit names make it clear which skill you want; automatic selection depends on the host. See four worked examples and QUICK_START.txt.

What this package does

The installable package contains instructions, supporting references, and interface metadata. It adds no MCP server, account connection, background hook, installer, or runtime script. Optional Python tools in the source repository validate and build releases; they do not run when using a skill.

Tools and readable attachments can deepen a review, but pasted text is enough to start. The skills use available evidence, distinguish a proposal from a finished result, and keep external actions inside the user's authorization. They do not restore another conversation's files or provide permanent memory.

These are instructions for an AI host. They cannot enforce a sandbox or guarantee correctness, confidentiality, immunity to prompt injection, or independent expert review. Keep secrets out of shared inputs and use the host's permission controls. For detailed boundaries and private vulnerability reporting, read SECURITY.md.

Contribute

Useful contributions improve real tasks with less effort for the user: clearer triggers, better examples, a demonstrated failure, or a smaller workflow. Please include a reproducible example with sensitive data removed. Start with CONTRIBUTING.md and our Code of Conduct.

Maintainers can validate and build with Python 3.11 or newer and Git. The Python tools use only the standard library:

python tools/validate.py
python -m unittest discover -s tests -v
python tools/build_release.py --output dist

Optional host checks require Claude Code:

claude plugin validate .claude-plugin/plugin.json --strict
claude plugin validate .claude-plugin/marketplace.json --strict

The CI checks package boundaries, references, metadata, license fingerprints, bounded secret signatures, and negative controls. Passing these checks is not a legal certification or a complete security audit. Changes to skills still need representative task checks.

License and sources

Published by YINGYE Studio. The package is licensed under Apache-2.0. Selected methods informed the newly written instructions; upstream licenses and attribution are retained in SOURCE_NOTICES.txt, licenses/, and provenance.json. Preserve them when redistributing. No upstream scripts or services are bundled.

See CHANGELOG.md for releases.