Your agent picks the right skill on its own.
Every skill says what it does and when to use it. ChatGPT, Claude Code, and Codex match those lines to your request and open the workflow that fits. You never have to remember its name.
A skill is a folder anchored by one instruction file.
The top of SKILL.md says what the skill does and when to use it. The host keeps that summary in mind and opens the rest only when the moment arrives. Some skills also include scripts, templates, or references that load only when needed.
Skillquiver/ ├── skills/ 24 shared │ ├── diagnose-systematically/ │ │ └── SKILL.md │ ├── verify-work/ │ ├── skillquiver-doctor/ │ └── … 21 more ├── .claude-plugin/plugin.json └── .codex-plugin/plugin.json
Skillquiver v2.2.0: 24-skill source candidate.
The update contains 24 shared Agent Skills, and the skills-only package works in ChatGPT and Codex. The same source catalog also supports Claude Code. It runs through host-approved tools, with no hosted backend, account, authentication, bundled MCP server, hooks, or app UI.
Version 2.2.0 is a local candidate for the existing listing under the public publisher name Drizzy07x. Upload, review, and publication remain separate external gates, so the current directory version may differ. See the privacy policy, terms, and support.
Skillquiver v2.2.0/ ├── writing-plans ├── diagnose-systematically ├── test-driven-development ├── improve-agent-instructions ├── skillquiver-doctor └── … 19 more
Choose your host.
The universal Plugins Directory is recommended for ChatGPT and Codex. The repository contains the 24-skill 2.2.0 candidate, while Claude Code can install that current source catalog directly from the repository marketplace.
Plugins Directory recommended
Open the Plugins Directory in ChatGPT, search for Skillquiver, note the displayed version, and select Add.
The new 24-skill catalog is available there only after the listing shows version 2.2.0. ChatGPT uses @ mentions for explicit skill selection. Repository installs are intended for local development and Claude Code.
Marketplace recommended
Run inside an interactive Claude Code session:
/plugin marketplace add Drizzy07x/Skillquiver
/plugin install skillquiver@skillquiverOr run the equivalent commands from a terminal:
claude plugin marketplace add Drizzy07x/Skillquiver
claude plugin install skillquiver@skillquiver --scope userUpdate with /plugin update skillquiver@skillquiver or claude plugin update skillquiver@skillquiver.
Manual
Copy the shared catalog for all 24 skills:
git clone https://github.com/Drizzy07x/Skillquiver.git
mkdir -p ~/.claude/skills
cp -r Skillquiver/skills/* ~/.claude/skills/Use a project's .claude/skills/ folder instead for a project-only install.
Marketplace recommended
Run from a terminal:
codex plugin marketplace add Drizzy07x/Skillquiver
codex plugin add skillquiver@skillquiverRefresh the marketplace and reinstall to update:
codex plugin marketplace upgrade skillquiver
codex plugin add skillquiver@skillquiverManual
Copy all 24 shared skills to the global agent directory:
git clone https://github.com/Drizzy07x/Skillquiver.git
mkdir -p ~/.codex/skills
cp -r Skillquiver/skills/* ~/.codex/skills/Use a project's .agents/skills/ folder instead for a project-only install.
v2.2 candidate: improve-agent-instructions adds scoped AGENTS.md and CLAUDE.md maintenance. Marketplace installs update normally; manual installations copy only skills/.
What each skill does, and when it steps in.
Every one of them runs in all three hosts.
Planning5
- brainstorming Asks what you actually want before a line of code is written. you're starting something new and the details aren't settled.
- writing-plans Turns the idea into a written list of tasks, in order. you know what to build but nobody has written down the steps.
- executing-plans Works through a written plan one task at a time, stopping only if it gets stuck. a plan already exists and you want it carried out.
- subagent-driven-development Hands each task to its own helper agent and reviews the result before moving on. the tasks in the plan don't depend on each other.
- dispatching-parallel-agents Sends out several helper agents at once, one per unrelated problem. you have two or more separate problems and no reason to fix them in order.
Execution7
- test-driven-development Writes the failing test first, then the smallest code that makes it pass. you're adding a feature or fixing a bug.
- solve-efficiently Finds the few files that actually matter instead of reading the whole project. the change touches several parts of the codebase.
- handle-host-boundaries Handles a skill or tool that belongs to another agent host without pretending it exists. a request names a capability from another host.
- execute-durably Keeps progress in a file on disk so long work can be picked up where it stopped. the job is long and might get interrupted.
- refactor-safely Tidies code without changing what it does, with tests in place first. code is messy or duplicated but still has to work exactly the same.
- using-git-worktrees Sets up a separate working copy so your current work stays untouched. you're starting a feature and don't want to disturb what's already open.
- communicate-clearly Keeps the answer as short as it can be without dropping the proof. you want less noise, or something explained in plain words.
Verification & debugging4
- diagnose-systematically Reproduces the failure, shrinks it, then tests one guess at a time. something breaks and nobody knows why yet.
- verification-before-completion Runs the command and reads the output before anyone says it works. you're about to hear “done”, “fixed”, or “all tests pass”.
- verify-work Checks finished work against real evidence and flags claims with nothing behind them. you want a second opinion on something that says it's complete.
- research-systematically Writes the question down first, then backs every answer with a source. comparing options, measuring, or checking how a library really behaves.
Code review3
- requesting-code-review Sends the finished work to a reviewer agent and collects what it finds. a task is done and about to be merged.
- receiving-code-review Checks each review comment against the code before changing anything. feedback arrives and some of it might be wrong.
- finishing-a-development-branch Runs the tests, asks how you want to land the work, then cleans up after it. the feature is finished and needs to go somewhere.
UI & automation2
- design-ui Commits to a look on purpose instead of the default one, builds it as one system, then checks it in a real browser. a screen looks generic, or you're redesigning one.
- automate-ui Clicks through a website or app for you and records what actually happened. something has to be tested, reproduced, or done in a real interface.
System2
- improve-agent-instructions Audits persistent agent guidance, moves each rule to the right scope, and keeps shared project instructions aligned. AGENTS.md or CLAUDE.md is stale, duplicated, contradictory, or in the wrong place.
- skillquiver-doctor Audits skills, plugins, and hooks for demonstrated conflicts, then offers reversible repairs one at a time. the wrong skill activates, a name shows up twice, or another plugin interferes.
One feature, from idea to merge.
These aren't 24 separate tricks. Almost every step of building something has a portable skill that fits it. A normal run in ChatGPT, Claude Code, or Codex looks like this:
- brainstorming Agree what you're building.
- writing-plans Write the steps down, in order.
- using-git-worktrees Get a clean workspace to build in.
- subagent-driven-development Build it task by task, reviewed as you go.
- test-driven-development Each task starts with a failing test.
- verification-before-completion Run it. Read the output. Then say it works.
- requesting-code-review Get the finished work reviewed.
- finishing-a-development-branch Merge it, then clean up.