Most marketers who try Claude Code for marketing start with the wrong skill. They install the repo, see the copywriting skill in the list, run it on their homepage, and get back confident-sounding output that's wrong about their product. They conclude the tool doesn't work. The tool works — they skipped the foundation.
[Insert diagram: a stack of marketing skills shown as boxes with arrows pointing down to a foundation block labeled "Product Marketing Context." A "Copywriting" box on top with a red arrow trying to skip the foundation, showing a "hallucinated output" warning.]
This is a focused setup guide for using Claude Code for marketing. It assumes you already understand what Claude Code Skills are and why they matter — if you don't, start with the broader Claude Code Skills explainer first and come back here when you're ready to install.
Here's the install path, the foundation skill that every other skill depends on, the order of operations that actually works, and the common mistakes that send people back to ChatGPT thinking the whole thing was hype.
Install Claude Code and the Marketing Skills repo
Three steps. About 20 minutes if you've never used a terminal before, 5 minutes if you have.
Step 1 — Install Claude Code. Follow the official Anthropic setup instructions: npm install -g @anthropic-ai/claude-code, then claude login. The login flow handles auth. You need a Claude account (Pro or Team works fine for solo use; the API rate limits aren't the bottleneck for most marketing work).
Step 2 — Install the Marketing Skills repo. From the skills.sh directory, run npx skills-cli add coreyhaines/marketing-skills. This pulls the skills into your local .claude/skills/ directory. Run it from inside whatever project directory you want the skills available in (or from your home directory if you want them globally available).
Step 3 — Verify. Open Claude Code in your project: claude. Then ask: "What marketing skills do I have available?" The agent should respond with the full list — Product Marketing Context, Customer Research, Positioning, Copywriting, SEO Audit, Programmatic SEO, and the rest. If it doesn't, the install didn't land — re-run step 2 from the right directory.
That's the install. No accounts, no API keys yet, no integrations configured. The skills are ready; they just don't have your context yet.
Run Product Marketing Context first — every time
The Marketing Skills repo has ~32 skills. Twenty-five of them reference one specific skill at the top of their workflow: Product Marketing Context. If that skill hasn't been run, every downstream skill either fails or hallucinates.
Product Marketing Context builds a structured profile of your product. It's a guided interview that captures:
- Positioning — category, ICP, differentiation
- Value propositions — top 3–5, ranked by audience priority
- Capabilities — what your product does, mapped to user outcomes
- Anti-features — what your product explicitly doesn't do (this is the part that prevents hallucination)
- Pricing — tiers, gates, expansion paths
- Competitors — direct, adjacent, status-quo alternatives
- Messaging hierarchy — H1-level claims, supporting proofs, objections handled
- Brand voice — tone rules, banned phrases, examples
The agent walks you through filling each section. Some answers it can draft from your website + GitHub README; the rest you confirm or correct. The whole thing takes 45–60 minutes the first time and produces a structured markdown file at .claude/product-marketing-context.md that every other skill reads.
Why this is non-negotiable. Without this file, the copywriting skill defaults to web search. It hits your homepage, maybe one blog post, maybe a review site from 2023. It writes copy based on whatever it scraped. The output sounds plausible because Claude is good at plausible — but the value props are wrong, the ICP is wrong, the differentiation is generic. You ship it, your prospects don't convert, you blame the AI.
With the file in place, the copywriting skill writes from accurate context. The SEO audit knows which keywords to skip because it knows what your product doesn't do. The email sequence skill writes from real value props instead of marketing-speak filler. This is the difference between AI marketing that works and AI marketing that's a dumpster fire — accurate product context is the foundation under everything.
Run this skill first. Update it whenever positioning shifts. Don't skip it.
[Insert screenshot: the Product Marketing Context skill in mid-run — Claude Code showing the structured prompt walking through value props, with the user confirming/correcting each one. Output shows the resulting product-marketing-context.md file structure on the right.]
The order of operations that actually works
After Product Marketing Context, there's a sequence. Skipping it is the second-most-common reason people get bad output.
Step 1 — Product Marketing Context. Foundation. Done above.
Step 2 — Customer Research. Mines existing reviews, interviews, support tickets, sales call notes. Feeds back into the Product Marketing Context file with real voice-of-customer language. If you have a Gong / Grain / Fathom library, point the skill at it. If you don't, the skill can scrape G2, Capterra, Trustpilot, and Reddit for review-mining.
Step 3 — Positioning (optional but recommended). Sharpens the category, ICP, and value proposition based on what Customer Research surfaced. Not always needed — if your positioning is already locked in, skip this and move to Step 4. If you're early-stage or pivoting, run this.
Step 4 — Pick the channel skill you actually need. Now you can run any downstream skill: Copywriting (homepage, landing pages, ads), SEO Audit, Programmatic SEO, Email Sequences, Ad Scripts, Brand Strategy, Web Design Review, whatever. They all read the foundation files and write from real context.
The phases — foundation, research, positioning, execution — are the same ones we recommend for SaaS marketing in general. The skills just make the phasing executable by an agent instead of a human team.
Common mistakes that send people back to ChatGPT
A short list of the failure modes we see most often:
Mistake 1 — Starting with the copywriting skill. Most common. Marketer installs the repo, opens it up, sees the copywriting skill, runs it on their homepage. Agent does a web search, writes generic-sounding copy, marketer concludes the tool is mid. Fix: always run Product Marketing Context first.
Mistake 2 — Asking the agent to "improve our marketing" without specifying. Too vague to trigger any specific skill. Claude Code defaults to general advice mode. Fix: be specific about the deliverable ("audit our SEO," "write hero copy for the new landing page," "build the welcome email sequence") and the agent will route to the right skill.
Mistake 3 — Not connecting integrations. Some skills (Programmatic SEO, Web Design Review, SEO Audit) work better when connected to specific tools — Ahrefs API for keyword data, Figma for design references, your CMS for direct writes. Without the integrations, the skills produce drafts instead of shipping work. Fix: set up the relevant integrations once. The Ahrefs key + a CMS write path covers 80% of the value.
Mistake 4 — Not iterating the SOP. A skill is markdown. If the output is consistently wrong in a specific way, edit the skill markdown to fix the SOP. Most users treat skills as static — they aren't. The whole point is that you can shape the agent's behavior by editing rules. Add a "don't do X" or "always check Y first" line, and the next run incorporates it.
Mistake 5 — Trying to run everything at once. New users sometimes install all 32 skills and try to use them in a single session. Claude Code's context window doesn't love this — too many skill files compete for trigger attention and the agent gets confused. Fix: run skills in sequence. Foundation, then research, then one or two channel skills at a time. Quit and restart the session if the context gets noisy.
What to try first (the 30-minute test)
If you've never run a marketing skill before and want to verify it actually works on your business, here's the test:
- Install Claude Code + the repo (5 minutes).
- Run Product Marketing Context end to end (30 minutes — the slow step).
- Then run the SEO Audit skill on your homepage (5 minutes). The audit reads your Product Marketing Context, looks at your live homepage, and produces a prioritized list of fixes — heading hierarchy, meta description quality, missing schema, internal linking gaps. We covered the underlying SEO issues in detail in H1, H2, H3 for SEO: Why Most SaaS Sites Get Heading Hierarchy Wrong — the audit catches all of those automatically and reads against your specific product context.
If the SEO Audit produces useful output (it almost always does), you'll have a concrete artifact to either ship or hand to a developer within 40 minutes of starting. That's the proof point that the tool actually works for you. Once you have that, the rest of the skills follow naturally.
Frequently asked questions
"Can I use Claude Code for marketing without using the Marketing Skills repo?"
Yes, but with much worse results. Out of the box, Claude Code has general capabilities and can do marketing-flavored tasks if you prompt it well. The Marketing Skills repo encodes the SOPs that make the agent reliably good at marketing work — without them, you're essentially using Claude Code as a smarter ChatGPT in the terminal. The repo is the differentiator. Install it; you'll see the gap immediately.
"Do I need to learn how each skill works internally?"
No — and this is the underrated point. You don't have to read the skill markdown to use it. Just ask Claude Code: "How do I use the SEO Audit skill?" and the agent will tell you what it needs, walk you through the inputs, and run the workflow. The skill itself becomes a tool the agent operates on your behalf. Reading the markdown is only necessary if you want to customize the SOP for your specific situation (which is the second-stage power move — adjust the rules to match how you think about the work).
"What if the skill produces wrong output for my product?"
Two possibilities. First, you skipped Product Marketing Context — the agent doesn't have accurate product info. Run that skill, then re-run the failing one. Second, the SOP in the skill doesn't match your situation. Edit the skill markdown to add the constraint that's missing ("always check X before doing Y," "never assume Z without confirming with the user"). The skill is text — you own it. Most users only realize they can edit the SOPs after they've been using the tool for a few weeks. Edit earlier.
"Are the Marketing Skills only useful for B2B SaaS?"
The skills were built with B2B SaaS in mind (because that's our domain at Conversion Factory), but the framework works for any product-marketing context: B2C, DTC ecommerce, agencies, infoproducts. The Product Marketing Context skill prompts for category-appropriate fields. The copywriting skill knows how to write for different audiences. The SEO skills are domain-agnostic. The main place B2B-bias shows up is in the example phrasing — the SOPs sometimes assume a sales-cycle pattern that doesn't apply to ecommerce. Easy fix: edit the skill markdown for your context.
"How often should I re-run Product Marketing Context?"
When something material changes — positioning shift, new pricing tier, ICP refinement, a competitor enters or exits the comparison set, a major capability ships. Otherwise it's a once-per-quarter check-in to validate that the file is still accurate. The foundation doesn't need constant rebuilding; it just needs to stay in sync with reality. If you're noticing your downstream skills producing slightly-off output, that's usually the signal that the foundation file is stale.
"Is this going to stop working when Claude Code changes?"
The skill format is stable. Claude Code's skill loading mechanism is documented and supported by Anthropic. The Marketing Skills repo is updated as the underlying tool evolves — we've shipped 1.1, 1.2, 1.3 over the last few weeks and will keep shipping. The skill markdown itself is portable; if a future tool (Manus, OpenClaude, Magister) consumes skills, the same files work there too. You're not locked into Claude Code specifically — you're investing in the SOP format, which is becoming a standard.
