How to build marketing AI skills for Claude Code and Codex
A marketing AI skill packages recurring instructions and supporting material so an AI can use them when needed. It can help you reuse the criteria you keep explaining for ad reviews, SEO outlines, or social drafts.
This guide shows how to choose a task, write a content-brief SKILL.md, and test it. Start with instructions in plain language.
Marketing workflows shared on GitHub
Corey Haines' Marketing Skills and Anthropic's Knowledge Work Plugins are examples of marketing procedures published on GitHub. The useful idea is to record your recurring work as a procedure you can check.
On September 11, 2026, OpenAI's GPT-6 Astra guidance advocated focused descriptions, references loaded when needed, and revisiting overly prescriptive procedures. Applied to marketing, this suggests keeping the audience, inputs, and deliverable clear rather than adding work instructions indiscriminately. OpenAI's article
The shared Agent Skills format uses a folder containing SKILL.md, with a name, description, instructions, and optional supporting files. Agent Skills specification
Publication does not guarantee results for your business. Test with your audience, product, and available sources.
To adopt an existing procedure, use GitHub marketing skill comparisons and trial steps to choose the advertising, SEO, or social deliverable you need.
Separate a prompt, a skill, reference material, and data access
The following table is a way to plan your setup.
| What you need | Suitable use | Marketing example |
|---|---|---|
| A prompt for this request | A single question or revision | Shorten ad copy without changing its meaning |
| A reusable skill | Consistent judgment, production, and checking | Review sources and prepare an outline and draft |
| Reference material | Shared business facts and language | Product conditions, brand terminology, verified examples |
| Tools or data connections | Current measurements | Retrieve ad spending or Search Console queries |
Writing a skill does not automatically provide account data or search demand. Arrange exports or connections separately. See how to give AI your marketing data.
Start with one deliverable you can assess
Choose work with an identifiable result rather than asking for every marketing activity to improve.
| Task | Material to provide | Initial deliverable |
|---|---|---|
| Ad review | Comparable reporting periods, conversion definitions, change history | A comparison and questions to investigate |
| SEO | Query, inspected search results, primary sources, existing articles | An outline answering the reader's question |
| Social content | Published source article, platform, audience, purpose | A draft preserving the source's meaning |
| Content production | Brief, product information, writing examples | A draft and unverified claims |
For an SEO skill, an outline ready for review can be the first completion condition. If publishing belongs in the eventual workflow, define the destination and checks separately.
Write the use case and procedure in SKILL.md
The following is this article's proposed starting template. It is not copied from a public skill or presented as an evaluated product. Adapt it to your materials.
---
name: marketing-content-brief
description: Prepare briefs for SEO articles or social posts when an outline is requested from an audience, goal, and verified sources. Do not use for publishing or changing ad settings.
---
# Prepare a content brief
## Input
Confirm the audience, reader's question, medium, goal,
verified sources, existing content, and intended next action.
Group necessary questions instead of guessing missing material.
## Procedure
1. Separate source facts from the writer's proposals.
2. For SEO, check search intent and overlap with existing articles.
3. For social posts, choose the source claim and destination.
4. Organize the opening answer, outline, examples, and next action.
5. Identify unverified specifications, figures, and achievements.
## Output
Show the purpose, reader's answer, outline, sources,
unresolved claims, and checks needed before drafting.
Do not invent search volume, testimonials, or customer counts.
Do not publish, send messages, or change advertising settings.
A description naming the deliverable and use case makes testing clearer than “act as an excellent marketer.” Before pasting extensive company information into the instructions, organize just the references the task needs. Anthropic's authoring guidance
If a reference contains prices or product specifications, include its owner and verification source so outdated prices do not become a repeated output.
Save it where Claude Code or Codex reads project skills
Use the location for your environment.
| Environment | Project location | Explicit test |
|---|---|---|
| Claude Code | .claude/skills/marketing-content-brief/SKILL.md |
Request an outline with /marketing-content-brief |
| Codex | .agents/skills/marketing-content-brief/SKILL.md |
Request an outline with $marketing-content-brief |
Claude Code documentation, OpenAI's skill guide
These examples cover local projects in Claude Code and Codex. Check the relevant loading and sharing methods for Claude chat or ChatGPT distribution. A shared format does not imply identical locations, tools, or permissions.
When references grow, put them in a references/ folder and explain when to read them. For multiple AI environments, keep one authoritative procedure and have each entry point read it to reduce divergence.
Test activation separately from work quality
Loading the file and producing a useful outline are different checks. These cases are suggested trials for your adaptation.
| Request | Expected behavior to inspect |
|---|---|
| Request an outline with the skill named | Uses sources and returns an outline with unresolved points |
| Request an outline without naming it | Selects the appropriate skill in a supporting environment |
| Request achievements without product evidence | Identifies missing information instead of inventing claims |
| Provide an article answering the same search intent | Avoids duplication and identifies useful additions |
| Request publication or advertising changes | Does not execute actions outside this production skill |
If a trial fails, check missing inputs, description scope, and output criteria before adding more instructions. Retest with the same material and compare review and correction effort.
Faster drafts or more text do not establish better ad performance or SEO rankings. Use the AI article prepublication checklist, then evaluate outcomes separately.
Align business facts with reusable work
Choose one recurring request and adapt the template to its inputs and deliverable. For a broader workflow, see marketing AI agent adoption and delegation boundaries.
Shoryaku AI chat uses registered business information to help discuss content outlines and drafts. Use it alongside your own Claude Code or Codex skills to organize product and audience facts and the procedures you repeat.