Yuzzyuk/marketing-os: A Claude Skill That Runs Fourteen Marketing Modules

The gap Marketing OS claims to fill

Most marketing prompt collections give advice. Marketing OS, according to its README, was assembled by tearing down the most-starred marketing skill repositories on GitHub and keeping the structures that worked while adding three things the author says every one of them lacked: scores on everything, artifacts instead of advice, and an explicit statement of what the skill could not determine.

The target user is a marketer or founder who already runs an agent host and wants structured output rather than a chat. Fourteen modules cover website audits, GEO (AI-search citability in ChatGPT, Perplexity and AI Overviews), copy variant generation, a hook engine, paid-ads fatigue diagnosis, email sequences, social posts, launch playbooks, positioning and pricing, competitive teardowns, app-store optimisation, and analytics. Two supporting directories, slop-patterns and rubrics/specs, are not standalone modules but run underneath the others: the AI-tell catalogue is applied to all prose the skill produces, and the rubrics define the scoring bands.

The honest framing is that this is a prompt architecture with a routing layer, not software. That distinction matters for anyone expecting a service to call.

How the router and progressive disclosure actually work

The README describes a progressive disclosure model. A router file loads first at roughly 2k tokens. When a task arrives, only the module matching that task loads. The stated consequence is that fourteen modules cost nothing until used, which is a token-budget argument rather than a capability argument.

Multi-dimensional work, such as audits, teardowns and variant generation, is described as subagent-native: the skill fans work out in parallel when the host agent supports subagents. On hosts that do not, the same work runs sequentially, so the wall-clock difference between Claude Code and a plain chat window is real but is a host property, not a skill property.

The repository is pure markdown with no executable code, which the README presents as a trust advantage: nothing to review before running it. The trade-off is the mirror image. There is no runtime to enforce a schema, so a malformed module output is a prompt problem you debug by editing markdown, not a stack trace.

One integration is documented beyond text output. If an ad-generation MCP such as Arcads is connected, the skill is stated to generate the briefed assets directly, turning the paid-ads path into diagnose, brief, produce within a single session. That is the only point in the README where the skill reaches outside the conversation.

Installing Marketing OS in Claude Code, Codex or Cursor

The README gives a two-command install for Claude Code through the plugin marketplace. Run them from a Claude Code session:

/plugin marketplace add Yuzzyuk/marketing-os
/plugin install marketing-os@marketing-os

The first command registers the repository as a marketplace source; the second installs the skill from it. In the web or desktop Claude client there is no command path. The README instead points to Settings, then Skills, then the plus button, and asks you to upload marketing-os.zip from the latest release, with code execution enabled in the account.

For Codex, Cursor or any other agent, installation is a file copy. The README's example destination is the Claude skills directory, but the principle is the same for any host that reads a skills folder:

cp -r skills/marketing-os/ ~/.claude/skills/marketing-os/

Before the first real task, the README insists on one setup step. Copy the template and fill it in:

cp brand-context.template.md brand-context.md

Place the filled file in your project root or in .claude/. Every module reads it. The README is blunt about the failure mode: skip this and the output is competent and interchangeable, which it calls the one thing marketing cannot be. That is the strongest claim in the document and also the easiest to verify, since the difference between a generic audit and one that names your actual offer should be visible in the first report.

The honesty spine and where it can still mislead

The README lists an honesty spine among the design principles: no invented proof, no fake precision, no winners declared on ten conversions, and every score labelled as the heuristic it is. The audit module outputs a weighted 0-100 with prioritised fixes, and the copy module generates 15 to 20 variants that pass through an expert-panel score before a de-slop pass.

Labelling a score as heuristic is the right instinct, but it does not make the score comparable across sites or over time. A weighted 0-100 built from a rubric file is only as stable as the rubric, and the README does not describe versioning of scoring bands between releases. If you track audit scores month over month and the rubric changes in a later version, the delta you report may be rubric drift rather than improvement. That is a real limitation for anyone planning to use the numeric output as a KPI.

The second constraint is coverage. The modules are text-generation and analysis tasks. There is no crawler, no analytics connector and no scheduled job. The analytics module is described as honest data reads and proper test design, which means it reasons about data you paste in; it does not fetch it. Anyone wanting automated weekly reporting is looking at the wrong tool.

Marketing OS compared with a general-purpose agent setup

The obvious alternative is not a competing marketing product but the default: a capable agent host with no skill installed and a well-written prompt per task. The difference in approach is structural rather than qualitative. A one-off prompt has no router, so the model sees only what you paste. Marketing OS loads a router and then a module, which means the scoring bands, the slop catalogue and the store field rules are present without you restating them each session.

That structure has a cost. You inherit the author's opinions about what a good hook is, which eighteen tactics matter, and how a 0-100 audit should be weighted. With your own prompts, those judgements are yours. For a team with a strong existing point of view on positioning or pricing, the positioning module may argue with your strategy rather than execute it, and the README offers no configuration layer for overriding module logic beyond editing the markdown directly.

A second alternative is a hosted marketing analytics or SEO platform. Those fetch data, schedule crawls and store history. Marketing OS does none of that. The two are complementary, not substitutes, and the skill's GEO and audit modules are most useful when fed real crawl or search data from elsewhere.

Licence, maintenance and upgrade cost

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the licence and copyright notice are retained. Two practical notes follow from that, offered as observations rather than legal advice. First, if you fork the skill and ship it inside a client deliverable, the MIT notice needs to travel with it. Second, because the modules are markdown you will edit, a fork will diverge from upstream quickly, and merging later releases becomes manual.

The last push was on 2026-08-17, the same date as the v1.1.0 release. The repository is not archived. Upgrading through the Claude Code plugin commands should pull the current version, but the README does not document rollback or pinning to a specific release, so a team that depends on stable audit scoring has no documented way to freeze a version short of copying the skills directory and maintaining it themselves.

Maintenance cost is mostly your own. There is no build step and no dependency tree, but the brand-context file needs updating whenever positioning, pricing or the offer changes, and every module reads it.

Editorial conclusion

Adopt Marketing OS if you already work inside Claude Code, Codex or Cursor and want scored, artifact-producing marketing output rather than another chat prompt. Skip it if you need a hosted dashboard, scheduled reporting or anything that runs without an agent host, since the repository contains no executable code. Before committing, read the audit module's rubric file to see how the 0-100 weighting is derived, and fill in brand-context.md first, because the README states that skipping it produces output that is competent and interchangeable.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. Yuzzyuk/marketing-os on GitHub