An open project for Claude Code and Codex
Claude-Codex Skills
A skills library where I turn practical lessons from working with AI agents into reusable guidance, from defining the task to verifying the result.
Why I built it
An AI agent can produce a lot of code quickly, but speed alone does not guarantee the right result. In my work, I repeatedly had to bring the focus back to the task outcome, the boundaries of the change and evidence that the solution actually worked.
I started turning those recurring requirements into standalone skills. Each skill gives the agent a concrete outcome, a way of working and completion criteria, without requiring an orchestrator or one universal workflow.
How I use the skills
I select one skill for the current task to define how the agent should work and when the task is complete. The repository helps me apply the same engineering discipline across codebases and stages of delivery.
Before development
I examine the opportunity, requirements, architectural constraints and plan before changes become expensive.
During implementation
I stay within the agreed scope, identify the root cause and where to fix it in the code, and require checks of the affected behavior.
When reviewing the result
I separate confirmed facts from assumptions. I state which checks found errors, which have not been completed and what risks remain.
During delivery and operations
I keep code readiness, publication and the observed state of the running system distinct.
My contribution
I designed the library structure, wrote and continue to refine the instructions, defined the responsibility boundaries between skills and established a shared approach to verifiable results.
I also prepared the Claude Code and Codex plugins, catalog integrity checks and behavioral validation scenarios. The repository is open so the skills can be inspected, adapted and used independently.
What practical use has taught me
A focused task works better
One skill with a clear result is easier to apply and assess than a large mandatory process.
Context matters more than instruction volume
A skill is useful when the agent first studies the actual code, constraints and existing decisions.
Verification must be visible
A list of actions is not evidence. A useful result shows what was verified and what remains unknown.
The engineer remains accountable
Skills help direct the agent's work, but choosing a solution and deciding whether it is sufficient still requires professional judgment.
Example: fix a language navigation bug
A worked example of a bounded task, not a performance benchmark. Assume that a link from a Spanish page opens an English page.
Use ln-41-surgical-change-implementer. Make internal navigation preserve the language from the URL. Verify direct links, reloads and Back/Forward navigation. Keep analytics events working. Do not publish the changes.
Review the resulting diff and the checks: the destination, displayed language and canonical URL should agree. Open the same link without JavaScript, reload it and navigate back. A passing build alone does not demonstrate that these scenarios work.
Read the skill instructionsUsing the skills
Does Claude-Codex Skills work with Claude Code and Codex?
The repository provides plugins for both tools. Follow the current installation instructions for your host; the skills guide the agent and do not replace its tools or permissions.
Do I need to run the entire collection?
No. Select a skill for the outcome you need. For a bounded code fix, Surgical Change Implementer covers implementation and its required checks; a separate review is useful when you need additional evidence.
What does a skill not guarantee?
A skill cannot guarantee correct code, access to a test environment or a successful deployment. Review the evidence and unresolved checks for your own project before accepting the result.
Documentation lives with the code
The current catalog, the purpose of each skill and installation instructions are maintained in the repository README.