Build a Codex SEO Skill Around Evidence, Scope, and Review.

Design the Codex skill as a bounded operating procedure

OpenAI documents a skill as a directory with a required SKILL.md and optional scripts, references, assets, and agent metadata. For SEO work, keep the main instructions short enough to follow, move detailed reference material into focused files, and include scripts only when a deterministic reusable command genuinely reduces errors.

Write the name and description for accurate triggering. The description should state that the skill handles scan-backed public website SEO implementation and should not trigger for generic marketing ideas, private application changes, or unsupported ranking predictions. Progressive disclosure makes this boundary important because the description is visible before the full instructions load.

Require concrete inputs: canonical production URL, affected route, observed behavior, expected behavior, scan evidence, public-versus-private boundary, known framework, and normal project checks. If the user supplies only improve SEO, the skill should collect evidence or identify a narrower task instead of editing broadly.

Begin the workflow by reading applicable AGENTS.md files completely. Then locate the public route, controller or entry point, shared renderer, content source, metadata logic, sitemap generation, robots policy, styles, assets, and relevant tests. Return an impact map before changing files so the user can verify the intended scope.

The implementation contract should require the smallest correct change, preserve user edits, avoid frameworks or dependencies unless authorized, and protect authentication, billing, APIs, webhooks, dashboards, analytics, secrets, and user data. Content must use verified product facts, and visible FAQs must match structured data exactly.

The verification contract should run file-specific syntax checks and tests, resolve new internal links, parse changed XML, inspect rendered metadata and schema, and use a real browser when layout or behavior changes. The skill should request a review of the final diff and remove generated test artifacts before handoff.

End with evidence boundaries. State what passed locally, what was deployed, whether the anonymous production route matches expectations, and whether the public rescan changed. Do not call rankings, AI citations, crawler demand, referral traffic, or conversions complete merely because the code and page render correctly.

Maintain the skill from real failures. Refine its description when triggering is wrong, add a reference when the same knowledge is repeatedly needed, and add a script only for stable deterministic work. Do not turn one skill into a framework that tries to cover every platform, CMS, deployment system, and content strategy.

Frequently asked questions

Where does Codex load repository skills from?

OpenAI documents repository skill discovery under .agents/skills from the current working directory toward the repository root. Verify the current documentation when structuring a shared skill.

Can I invoke a Codex skill manually?

Yes. OpenAI documents explicit invocation through the skill selector or mention syntax, while Codex can also select a skill when the task matches its description.

Can the skill enforce every safety rule?

No. Instructions guide agent behavior, while permissions and repository controls provide stronger enforcement for sensitive actions.

Related internal links