Marketing Skills
Marketing Skills architecture: context, procedures, references and evaluation data
Map the library’s document architecture and distinguish host-driven execution from repository maintenance automation.
What you will learn
- The host agent supplies execution and permissions.
- Synchronization maintains metadata, not campaigns.
- Evaluation prompts are not test-run evidence.
Before you start
- Basic Markdown and repository navigation
- An owned page and a factual product brief
Explain the chapter’s workflow boundary and check its evidence or proposed acceptance task.
Key takeaways
- The host agent supplies execution and permissions.
- Synchronization maintains metadata, not campaigns.
- Evaluation prompts are not test-run evidence.
The runtime is outside the repository’s prose
A skill supplies a name, a triggering description and a body of instructions. The host agent decides how to load and follow that material within its own tool and permission system. A reference from one skill to another is a procedural dependency, not evidence that a scheduler or deterministic workflow engine executed a call graph.
Shared product context sits upstream of many tasks. References add detailed guidance, and optional integrations describe how external systems can supply data or accept changes. This architecture is flexible because it does not require a single analytics vendor, but reproducibility needs the host model, prompt, context revision and retrieved evidence as well as the skill revision.
Maintenance automation has a narrower job
.github/scripts/sync-skills.js enumerates direct skill directories containing SKILL.md, reads metadata, sorts names and regenerates a README table. It updates the first marketplace plugin’s count and removes a stale explicit skills array there. Separately, it aligns plugin.json version with marketplace metadata.version when that value exists.
The inspected sync workflow runs on selected main-branch changes and commits a restricted file pattern containing the marketplace, plugin and README. This is catalogue maintenance, not content publication or campaign execution. The distinction matters when someone sees an automated commit and assumes their marketing deliverables were also generated or deployed.
Evaluations are specifications, not observed outcomes
The ai-seo evaluation file contains prompts, expected outputs and assertions. It describes desired behavior; it is not a log of executed model responses. A structural skill-validation workflow also exists, but its presence does not establish that every claim in every reference is current or that a real reader learned the intended concept.
An evaluation can preserve an old assumption after the prose changes. In the inspected files, some expected outputs emphasize fixed citation statistics while newer instruction text discusses format volatility. Review the assertions alongside the skill and its sources. Passing a stale rubric can reward precisely the statement your current editorial policy should question.
Implementation steps
- 1
Map context, skill, references and integrations separately.
- 2
Record the host and evidence used for a task.
- 3
Compare evaluation assertions with current skill wording.
Copy-ready example
Product context -> selected skill -> relevant references
Host agent + permitted tools -> proposed artifact -> review
Maintenance: skill folders -> metadata -> README / plugin catalogue
Evaluation prompts are a separate specification.Frequently asked questions
Is a skill reference a deterministic function call?
No. It is an instruction to the host agent; actual execution depends on that host and the task context.
Does the evaluation JSON prove tests passed?
No. Expected outputs and assertions need a separate run with retained responses and review evidence.
Sources
- README.mdSource checked 2026-09-08
- LICENSESource checked 2026-09-08
- .claude-plugin/plugin.jsonSource checked 2026-09-08
- .github/scripts/sync-skills.jsSource checked 2026-09-08
- .github/workflows/sync-skills.ymlSource checked 2026-09-08
- .github/workflows/validate-skill.ymlSource checked 2026-09-08
- scripts/sync-partners.mjsSource checked 2026-09-08
- skills/product-marketing/SKILL.mdSource checked 2026-09-08
- skills/ai-seo/SKILL.mdSource checked 2026-09-08
- skills/seo-audit/SKILL.mdSource checked 2026-09-08
- skills/programmatic-seo/SKILL.mdSource checked 2026-09-08
- skills/ai-seo/evals/evals.jsonSource checked 2026-09-08
- tools/PARTNERS.mdSource checked 2026-09-08