Agent skills
Personal agent skills for engineering, review, investigation, and verification.
Repository structure
Keep skills/ as the shared source for all agents and environments. Individual
skill installers such as npx skills use those folders directly. The portable
plugin.json, Codex compatibility manifest in .codex-plugin/, and marketplace
catalog in .agents/plugins/ add distribution formats without moving or copying
the skill sources.
A future Claude Code plugin can add its own manifest and marketplace metadata
pointing to the same skills/ directory. Keep host-specific packaging outside
the skills, preserve each skill's existing invocation metadata, and avoid forks
or generated copies of skill instructions for each host. This change packages
Codex distribution; it does not replace individual skill installation or add a
Claude Code plugin installer.
Install as a Codex plugin
This repository is also the plugin: inprealpha-agent-skills, in the inprealpha
marketplace. The existing skills/ directory is the single source of truth; plugin
packaging does not change skill behavior, licenses, or invocation policies.
In a current Codex CLI:
codex plugin marketplace add inprealpha/agent-skills --ref main
codex plugin add inprealpha-agent-skills@inprealpha
Start a new chat after installation. In the desktop Plugins Directory, select the
inprealpha source and install Abir's Agent Skills if you prefer the UI.
To refresh it later, run codex plugin marketplace upgrade inprealpha
and start a new chat.
Use in Codex Cloud
Installing a local marketplace does not by itself provision cloud sessions. Choose the route your account supports:
Workspace plugin: A workspace admin can open Admin > Plugins > Add > Import
marketplace, enter https://github.com/inprealpha/agent-skills, leave Path empty,
and set Branch to main. Import and make the plugin
available to your role. Install/enable it, then start a new cloud task and check
that its skills are available. The bundle contains no MCP servers or desktop hooks.
Cloud project setup: If workspace import is unavailable, ask the environment
setup agent to install the same skill folders into your target project's
.agents/skills/. Add this repository to the environment, then run:
python3 /path/to/agent-skills/scripts/install_cloud_skills.py --project /path/to/your-project
Replace those paths with the actual cloud checkout paths. Alternatively, from the target project root, use this setup snippet (requires Python 3, Git, and GitHub network access):
skills_checkout="$(mktemp -d)"
git clone --depth 1 --branch main https://github.com/inprealpha/agent-skills.git "$skills_checkout" &&
python3 "$skills_checkout/scripts/install_cloud_skills.py" --project "$PWD"
Publish/republish the prepared environment and start a new task in that project. For reproducible environments, use a reviewed commit rather than following the moving branch. The helper copies every supporting file and license, accepts an identical repeat install, and refuses differing existing skills before copying anything. It does not overwrite or remove project skills; review conflicts when upgrading. This is repository skill discovery, not a cloud marketplace install.
Invoke and verify
Select a skill with @ where the UI supports it, or $skill-name in Codex CLI.
For a first check, ask: “Use the architect skill to propose a design for this
change; read its SKILL.md first and keep this request design-only.” Check that the
agent actually reads the installed skill. Ten skills preserve explicit-only
invocation through agents/openai.yaml; craft, orchestrate-work, review-loop,
and simplify-code retain automatic discovery.
The plugin supplies instructions, not capabilities: delegation, browser control,
connected services, and project runtimes must be available in the cloud session.
It does not include personal config.toml, credentials, conversation history, or
global AGENTS.md. Avoid installing duplicate standalone and plugin copies in the
same session. Cloud import and runtime behavior must be verified in your workspace.
Official references: plugin packaging, workspace GitHub import, skill discovery, and cloud environments.
Validate and build an uploadable ZIP
Run these commands from a repository checkout. Development dependency: Python
3.10+ and PyYAML==6.0.2. The cloud setup helper uses
only the Python standard library. CI validates packaging, tests the setup helper,
and publishes a ZIP artifact on each push or pull request.
python3 -m pip install PyYAML==6.0.2
python3 scripts/package_plugin.py --output dist/inprealpha-agent-skills.zip
python3 -m unittest discover -s tests -v
The ZIP puts manifests and skills/ at its root, includes the cloud setup helper,
and excludes Git history, build tooling, CI files, and marketplace metadata.
Existing per-skill license notices are retained; this
package does not grant a new repository-wide license.
Install individual skills
Use the skills CLI:
npx skills add inprealpha/agent-skills
List the available skills without installing:
npx skills add inprealpha/agent-skills --list
Install one skill:
npx skills add inprealpha/agent-skills --skill simplify-code
Skills
| Skill | Description |
|---|---|
| architect | Design caller usage, interfaces, data types, and module ownership before implementing a change. |
| blast-radius | Assess what a change could break beyond its diff and test the assumptions its safety depends on. |
| craft | Simplify writing, documentation, code, architecture, interfaces, and creative work so people can quickly understand and use it. Use for “AI slop,” confusing or overcomplicated output, project reviews, and making the current result ready to share or ship. Not an AI detector. |
| create-verification-skill | Create executable project verification instructions with a feature map and retained evidence. |
| hillclimb | Improve one measurable outcome through isolated hypotheses, repeated measurements, and regression checks. |
| how | Explain a subsystem through traced runtime flow, data ownership, and concrete code references. |
| interrogate | Adversarially review a change and deliver a prioritized verdict grounded in reachable failures. |
| maintain-verification-skill | Audit a project verification skill against current source and observed application behavior. |
| orchestrate-work | Coordinate substantial work with ownership suited to the task, parallel execution, and fresh perspectives. Use for broad ambiguous goals, separable workstreams, or requested team workflows. |
| reflect | Extract durable lessons from a work session and route them to focused skill or tooling improvements. |
| review-loop | Finish work through repeated fresh-context reviews and fixes. Use when the user wants an autonomous review loop after a brief alignment on the outcome and review perspectives. |
| simplify-code | Simplify code when asked to review over-engineering or apply cleanup, including redundant tests and avoidable dependencies. Preserve behavior; keep unrelated feature work and general correctness reviews outside this skill's scope. |
| skill-evaluation | Compare skill variants on realistic tasks with controlled inputs and evidence-based scoring. |
| why | Investigate the history and rationale behind code, with cited facts and explicit uncertainty. |
Skill instructions and required supporting files are in skills/. Harness support depends on the tools available in the current agent environment; cross-harness behavior has not been comprehensively tested.
Attribution
Adapted upstream material retains its MIT license notices within each skill directory:
architect,blast-radius,create-verification-skill,hillclimb,how,interrogate,maintain-verification-skill,reflect,skill-evaluation, andwhy: pstack by Lauren Tan, revisionf5bdd6826fd0a0d9cbc4347134c3a74a200b9d9d.orchestrate-workandreview-loop: selected procedures and writing guidance from Matt Pocock skills, revision3cca18b368ae95cdbdebbff572ccafa662551015.simplify-code: Ponytail by DietrichGebert, revision356918eba965ee1eac64bd3a7f0dd02108350de5.
craft is original work and includes a notice preserving the existing policy: no additional license is granted. No additional repository-wide license is specified for original contributions.