Marketing Skills for Claude Code

12 AI marketing skills in one Claude Code plugin: competitor research, social media content calendars, LinkedIn and X posts, SEO content briefs, ad copy, email sequences, personalized cold email outreach, TikTok / Reels / YouTube Shorts scripts, Reddit and Hacker News posts, and approval-gated publishing to n8n, Make, or platform APIs.

claude plugin marketplace add marcosmodly/marketing-skill
claude plugin install marketing-skill@marketing-skill

Every skill is a standard SKILL.md in skills/, shares one brand-voice config, and works on its own or chained together. Nothing posts or sends without your explicit approval. Using OpenAI Codex? See Using with Codex.

If it saves you time, a star on the repo helps other people find it. Made something with it? Share it in Show and tell.

See it

31 video templates, rendered on your machine with music composed in your brand’s own sound. Each frame below is one template, mid-video:

Marketing videos (ads, demos, narrated explainers, testimonials, FAQs, launches, team)

Everyday posts: meme captions, comment bait, skits, and stories

3D and pixel-art animation, narrated, with captions synced to the voice

The plugin’s own explainer was made with the same renderer.

What you can do with it

• Competitor analysis: research a rival and get a SWOT, messaging breakdown, and sales battlecard with cited sources.
• Social media content calendar: batch-write a week or month of dated posts for LinkedIn, Twitter/X, newsletters, Reddit, and short-form video, without repeating recent topics.
• Content repurposing: turn one blog post or announcement into a LinkedIn post, an X thread, and a newsletter blurb.
• SEO content briefs: target keyword, search intent, related keywords, outline, and meta title and description, with no made-up search volumes.
• Ad copy: A/B-testable variants for Facebook and Instagram (Meta), Google Search (RSA), and LinkedIn Ads, sized to each platform’s limits.
• Email marketing: drip, nurture, and welcome sequences with timing, subject lines, and preview text.
• Cold email outreach: a daily batch of researched, 1:1 personalized prospecting emails, deduplicated against a permanent contact log that syncs replies, bounces, and unsubscribes from Gmail.
• Short-form video: hook-first scripts plus caption, hashtags, and best time to post for TikTok, Instagram Reels, and YouTube Shorts, shaped to the right type of marketing video (how-to, testimonial, launch, FAQ, and more) or to an everyday FYP format (POV, tier list, text-message skit, storytime, hot take, and 17 more). Also the finished vertical video itself, rendered on your machine: animated text over stock or AI backgrounds, or a narrated 3D or pixel-art animation, with a voiceover, music composed in your brand’s own sound, and meme sound effects.
• Community posts: Reddit, Product Hunt, Hacker News (Show HN), Indie Hackers, dev.to, Discord, Slack, and Telegram posts written to each community’s actual rules, with a Go/No-Go before drafting.
• Visual briefs: shot lists, per-scene image/video prompts, and aspect ratios for any AI image or video generator.
• Publishing: hand finished content to n8n, Make, or any webhook, or post directly to LinkedIn, X, Facebook, Reddit, Discord, Slack, Telegram, dev.to, YouTube, Instagram, or TikTok with your own API credentials.

See a full worked run → — one continuous full-pipeline call from research to the approval checkpoint before anything actually publishes.

Skills

Plus one setup command: /marketing-skill:marketing-setup.

Using your own project as source material

competitor-research, content-calendar, content-repurposer, seo-brief, ad-copy-generator, email-sequence, email-outreach, visual-brief-generator, short-form-video, and community-post-generator can pull from the project they’re installed in instead of requiring you to paste content every time. If you reference “our product,” “our feature,” “our changelog,” etc. without providing the text, they’ll check README*, CHANGELOG*, docs/**/*.md, and package.json/pyproject.toml at the project root first, and ask you directly only if nothing relevant turns up. They only read documentation-oriented files this way — never arbitrary source code — so install this in your product’s own repo to get the most out of it.

Install

Plugin method (recommended):

claude plugin marketplace add marcosmodly/marketing-skill
claude plugin install marketing-skill@marketing-skill

This works once the plugin’s branch is merged to the repo’s default branch, since marketplace add tracks a repo’s default branch. To try it before merging (e.g. against a feature branch), clone the repo locally and point at your checkout instead — this is verified to work regardless of which branch is checked out:

git clone https://github.com/marcosmodly/marketing-skill.git
cd marketing-skill && git checkout    # if not already on it
claude plugin marketplace add ./
claude plugin install marketing-skill@marketing-skill

Manual fallback: no confirmed minimum Claude Code version exists for the plugin system — if claude plugin isn’t available on your install (check with claude plugin --help), copy the skills in by hand instead:

cp -r skills/* ~/.claude/skills/        # personal, all projects
# or
cp -r skills/* /your/project/.claude/skills/   # project-scoped

The manual copy skips the setup command (see “Configure” below for the equivalent manual step) and ${CLAUDE_PLUGIN_ROOT} path substitution. Copied skills can’t find references/, state/, or scripts/ on their own, so set CLAUDE_PLUGIN_ROOT to your clone of this repo before starting Claude Code (export CLAUDE_PLUGIN_ROOT=/path/to/marketing-skill); otherwise each skill asks you where the plugin lives.

Using with Codex

OpenAI’s Codex reads Claude Code’s plugin files (.claude-plugin/), so the same repo installs there too:

codex plugin marketplace add marcosmodly/marketing-skill
codex plugin add marketing-skill@marketing-skill

To try a branch, add --ref to the first command, or pass the path to a local clone instead of marcosmodly/marketing-skill.

All 12 skills load as-is; Codex ignores the allowed-tools line. A few things differ from Claude Code:

• Setup command: Codex turns /marketing-skill:marketing-setup into a skill, so ask Codex to “run the marketing setup” instead of typing the slash command. Codex only converts commands under 4 KB, so keep commands/marketing-setup.md short when editing it.

• Plugin folder: Codex doesn’t fill in ${CLAUDE_PLUGIN_ROOT}. Each skill works out the plugin folder itself (from the CLAUDE_PLUGIN_ROOT environment variable, or by finding the folder that holds references/brand-voice.md). Setting the variable is the reliable option, and it keeps your brand voice and queued posts in a folder you control rather than in Codex’s installed copy:

  export CLAUDE_PLUGIN_ROOT=/path/to/marketing-skill   # your clone
  codex
  

• Sandbox: in its usual workspace-write mode, Codex only writes inside the folder you started it in and blocks network access from shell commands. If the plugin folder is somewhere else, add it to writable_roots (or approve the writes when asked). Publishing (publish_webhook.py, publish_direct.py) needs network access. Both settings go in ~/.codex/config.toml:

  [sandbox_workspace_write]
  writable_roots = ["/path/to/marketing-skill"]
  network_access = true
  

• Live research: competitor-research, seo-brief, community-post-generator, content-calendar, email-outreach, short-form-video, and full-pipeline research the web as they run. Set web_search = "live" in ~/.codex/config.toml so results are current rather than cached.

• Connected tools: Gmail drafts, prospecting tools, and video tools (Higgsfield, Figma Weave, Canva) work once the matching MCP server is set up in Codex. Without one, each skill falls back the same way it does in Claude Code.

Configure

Every skill reads references/brand-voice.md before producing output — your task priority, content types (short-form video, long-form video, text posts, email, images), target audience, tone, output format, banned words, and formatting constraints, defined once and reused everywhere.

• Guided setup: run /marketing-skill:marketing-setup any time (first-run or to update your answers).
• Automatic: if you skip setup, the first skill you actually use will ask the same 4 questions itself before proceeding, and save your answers for next time.
• Manual: edit references/brand-voice.md directly — keep its section headings intact so every skill can still find them.

email-outreach additionally reads its own references/outreach-strategy.md — ICP/target segment(s), value proposition and angle per segment, offer and primary CTA, daily volume, follow-up cadence, and suppression window. It’s configured the same way: guided setup the first time email-outreach runs, or hand-edit directly. Unlike brand-voice.md, it also re-opens itself automatically — the skill checks its “last refreshed” date on every run and walks you through a strategy-refresh conversation once 30 days have passed, using state/outreach-log.md’s accumulated send/reply/bounce counts to inform it.

Making videos

short-form-video writes the script, captions, hashtags, and posting times for TikTok, Reels, and Shorts, and, if you want, makes the video:

• From a template: 7 marketing types and 22 everyday formats (above), rewritten for your product and brand voice, over stock or AI backgrounds.
• Narrated: a script of numbered lines becomes a voiceover (a free local voice, or your own takes) over a background per line, with word-by-word captions synced to the voice (compose.js).
• Animated, in 3D or pixel art: the same narrated script, drawn as 3D scenes in your brand’s colors (your screenshot on a phone, your logo on a coin) or as a pixel-art game with chiptune music (style: 3d or style: pixel).
• Your own footage: a talking-head take from your phone gets its pauses cut and captions added (cut.js); screen recordings play in a phone frame.

Every video gets your sonic identity (one genre, key, and hook, so your sound becomes recognizable) and visual identity (colors, font, and a logo end card that lands with the sonic logo), the hook on screen from the first frame, and a QA pass for text the platforms’ buttons would cover or that leaves too fast to read. Hook variants render several openings to test, and a video log steers the next videos by how yours actually did. Setup is free: Node, Playwright, and ffmpeg.

cd scripts/video && npm install
node compose.js script.md                    # a narrated video page from a script
node render.js page.html out.mp4 --cover 0   # the video, its captions (.srt), and its cover

Everything else, from backgrounds and sound to templates and connected video tools: docs/video.md.

Content calendar & approval model

content-calendar batch-generates posts and queues them in state/content-calendar.md — a plain Markdown table that’s both the forward calendar and the permanent send history (rows accumulate; nothing gets cleared out). It’s safe to hand-edit directly. The table itself is just an index: each row’s Notes column points at the file under state/posts/ that actually holds that post’s full content — publish-pipeline reads the file, not the table, when it sends.

The file enforces a separation of duties that every skill in this plugin respects: generation and approval are always two different steps. content-calendar (and anything else that drafts content) can only write Planned, Drafted, or Ready for Approval to a row’s Status — never Approved. Only publish-pipeline writes Approved, and only immediately after showing the real content to a human in a live conversation and getting an actual reply back — silence, a timeout, or the request having been phrased as blanket authorization (“just post these”) never counts as that reply. publish-pipeline then flips the row to Sent right after a successful send.

When more than one row is ready at once, publish-pipeline lists them together as a numbered batch (each with its Account/Destination/ Attachments/Final Text spelled out) instead of asking “send this?” one row at a time — a reply has to name specific numbers to approve them, since an unqualified “send them all” invites rubber-stamping a list nobody actually read. It also flags any row more than 14 days old as Stale right in that listing, and content-calendar separately reports how many rows across the whole calendar have crossed that age every time it runs — a queue that’s easy to lose track of is exactly what turns a careful approval step into a bulk-approved backlog later.

This matters most the moment you try to run this plugin unattended — see “Running this on a schedule” below.

Running this on a schedule

Nothing in this plugin runs on a timer by itself — a skill only executes when a session invokes it. To get daily/weekly output, schedule the session, not the plugin: Claude Cowork’s scheduled tasks (or Claude Code’s own Routines/triggers) can fire a session on a cadence that runs content-calendar or full-pipeline.

That solves “run it every day.” It does not turn this into unattended auto-posting, on purpose: content-calendar only ever queues rows as Drafted/Ready for Approval, and publish-pipeline only sends a row already Approved by an actual human reply in that conversation — a scheduled trigger firing with nobody there to reply just leaves content queued for you to review later, which is the intended behavior, not a bug to work around. If you deliberately want unattended sending, you’d need to change that approval step yourself — this plugin doesn’t ship a way to do that, on the theory that a live company account posting unsupervised is a decision only you should make explicitly, not one a scheduling tool should make for you by default.

email-outreach follows the exact same model, just applied to cold email instead of social posts: schedule a daily session to get a consistent daily batch (its own “consistent daily outreach” requirement), and it queues every draft as Drafted/Ready for Approval — never Sent — exactly like content-calendar does, for the same reason. Its monthly strategy refresh doesn’t need a separate schedule at all: the skill checks references/outreach-strategy.md’s own last-refreshed date on every run and walks through the refresh conversation itself once 30 days have passed, so the one daily trigger alone covers both cadences.

Publishing, community posts, and outreach

• Publishing: hand finished content to n8n, Make, or any webhook (scripts/publish_webhook.py), or post straight to LinkedIn, X, Facebook, Reddit, Discord, Slack, Telegram, dev.to, YouTube, Instagram, or TikTok with your own credentials (scripts/publish_direct.py). Both refuse to run without --dry-run or --confirmed. The page covers what each platform’s access really takes, including TikTok’s private-only rule until your app is audited.
• Community posts: how community-post-generator researches each subreddit’s, Product Hunt’s, Hacker News’s, Indie Hackers’, dev.to’s, or a Telegram channel’s actual rules before drafting, and asks you for Discord’s and Slack’s.
• Cold outreach: the contact log that keeps anyone who replied, bounced, or unsubscribed from ever being emailed again, and how Gmail and prospecting tools plug in.

FAQ

What are Claude Code skills?

A skill is a folder with a SKILL.md file that teaches Claude a repeatable task. Claude reads each skill’s description and uses it automatically when your request matches (“write an SEO brief for…”, “give me a week of LinkedIn posts”), or you can call one by name, such as /marketing-skill:seo-brief.

Do I need API keys or paid tools?

No. Research uses Claude’s built-in web search and fetch, and every skill produces its full written output with no connector or API key. Gmail, a prospecting tool, Higgsfield, Canva, and platform API credentials are optional extras, and paid ones are used only when you ask. Rendering a video locally needs only free tools (Node, Playwright, and ffmpeg), plus free Pexels and Pixabay keys if you want stock backgrounds. Actually sending content needs your own webhook or credentials.

Will it post or send emails automatically?

No. Drafts are queued in state/, and only publish-pipeline sends, after you approve specific items in a live conversation. email-outreach creates drafts and never sends. See Content calendar & approval model.

Can I use just one skill?

Yes. Ask for what you need and only the matching skill runs. To skip the plugin system, copy the skill folders in by hand (see Install).

Repo layout

.claude-plugin/
  plugin.json         # plugin metadata
  marketplace.json     # lets this repo install itself via `marketplace add` (Claude Code and Codex)
skills/                 # the 12 skills, one SKILL.md each
commands/
  marketing-setup.md    # the /marketing-skill:marketing-setup command
references/
  brand-voice.md         # shared config every skill reads (hand-edited)
  outreach-strategy.md   # email-outreach's own config: ICP, angle, offer, cadence (hand-edited or guided setup)
  video-types.md         # 23 marketing video types: goal, short-form beats, and how each gets made (hand-editable)
  sound-guide.md         # music genres, sound effects, how to pick a project's sonic identity, a hook library, licensed tracks
  hooks.md               # hook patterns by goal, written to work with the sound off, and how to test them
  video-rendering.md     # the local render procedure short-form-video follows
  fyp-formats.md         # 22 everyday / FYP formats: why each works, beats, copy at two tones, weekly mix and rotation
docs/                    # the longer guides: video, publishing, community posts, outreach, and the template gallery
state/
  content-calendar.md    # calendar/history index (generated + appended to, hand-editable)
  posts/                  # one file per queued post's full content, linked from the index above
  outreach-log.md         # outreach queue/history + permanent suppression list (generated + appended to, hand-editable)
  outreach/               # one file per drafted email's full content, linked from the log above
  video-log.md            # every posted video with its format, hook, and numbers; steers the next ones
scripts/
  publish_webhook.py     # stdlib-only webhook sender (see --help)
  publish_direct.py      # stdlib-only direct-to-platform scaffold (LinkedIn/X/Meta/Reddit/Discord/Slack/Telegram/dev.to/YouTube/Instagram/TikTok), needs your own API credentials (see --help)
  video/
    render.js            # animated HTML page -> vertical MP4 with an original soundtrack (needs Node, Playwright, ffmpeg)
    soundtrack.js        # the music and sound-effects synthesizer render.js uses (no dependencies)
    media.js             # finds stock photos and clips (Pexels, Pixabay, Openverse) and writes their credits
    check.js             # QA pass over templates: layout lint, cues, fonts, loudness, contact sheets
    voice.js             # free local voiceover (Kokoro): a script -> one WAV per line
    captions.js          # word timing for a voice line, caption chunks, and .srt files
    compose.js           # a narrated video page from a script: voice, backgrounds (photos, 3D, or pixel art), synced captions
    brand.js             # applies the Visual Identity to a page: colors, font, logo end card
    cut.js               # a filmed talking-head take: pauses cut, captions, and a page to render it
    beats.js             # a music track's tempo, beats, bars, and drops, to cut a video on its beat
    gallery.js           # rebuilds the template gallery in docs/gallery
    templates/           # 7 marketing + 22 everyday + 3D and pixel-art starting pages, built by templates/build.py

Share what you made

Made a content calendar, a cold email batch, a video, or a launch post with these skills? Post it in Show and tell, along with which skills you ran and anything that didn’t work the way you expected. Questions go in Q&A, and bugs in Issues.

Contributors

• marcosmodly