Obsidian/Claude/How To Build One Claude Skill That Replaces Your Entire Marketing Stack In A Weekend.md at main

admin/Obsidian

Fork 0

Files

T

coco 1b8027e61c 时地方

2026-07-21 14:53:24 +08:00

16 KiB

Raw Permalink Blame History

title, source, author, published, created, description, tags

title

source

author

published

created

description

The four tools, the real prices, the real overlap

Let me show you the receipts before the build.

Notion AI sits in the Business plan at $20 per user per month on annual billing, $24 month-to-month. Anthropic's pricing page confirms it. The $10 standalone AI add-on was killed in May 2025, so if you want AI Agents and Ask Notion you have to jump from Plus to Business. For a solo founder, that's $240 a year minimum just to get AI inside Notion.

Jasper Pro is $59 per month annual, $69 monthly. Their pricing page is unchanged through April 2026. Brand voices, campaigns, templates. $708 a year if you stay on annual.

Buffer is $5 per channel per month on annual Essentials, $6 monthly. Connect five channels (X, LinkedIn, Threads, Instagram, Bluesky) and you're at $25 a month. $300 a year. Add a teammate and you jump to the Team plan at $10 per channel.

Hypefury Creator is $59 a month, or $590 annual. Engagement builder, autoplugs, evergreen recycling, Tweet-to-Reels. The Starter plan at $29 caps you at one X account and one month of scheduled content, which most active creators burn through in a single batch session.

Add it up: $20 + $59 + $25 + $59 = $163 a month on the annual cadence. Call it $158 if you skipped the Hypefury Creator tier and used Starter for the first month.

Now look at what these four tools actually do.

Generate marketing copy in your voice. Schedule that copy across platforms. Recycle the best-performing pieces. Brainstorm hooks. Adapt one piece of content for X, LinkedIn, and a newsletter. Track what worked.

That's it. Four logins, one job: take your ideas and turn them into a consistent, on-brand publishing engine.

A Claude Skill can do every step of that workflow. The scheduling part still needs an MCP server (we'll get to that). But the brain of the stack, the part you're paying $1,896 a year for, is a markdown file.

Why most Skills don't replace anything

Before the build, the contrarian frame.

Open any "best Claude Skills" article from the last three months and you'll see the same advice: install Corey Haines' marketing skills (6,852 GitHub stars). Install wondelai's strategy skills. Run npx skills add until you have 160 of them.

That advice isn't wrong. Those skills are excellent. But it misses what makes a Skill replace a SaaS subscription instead of just adding another tab to your workflow.

A community Skill is generic by design. It has to be, or it wouldn't help anyone. A Skill that replaces your stack is the opposite: it knows your voice, your offers, your audience, your formatting rules, and your scheduling cadence. It's the brand brief you'd give to a junior marketer, except it lives in a markdown file Claude reads automatically every time you ask for content.

Here's the test. If your SKILL.md could be lifted out of your folder and pasted into a stranger's setup with no edits, it's a tool. If pasting it into a stranger's setup would produce content that sounds like you, it's not really yours.

The leverage isn't in installing more Skills. It's in building one that's so specific to your business that it replaces the four generic tools doing the same job badly.

That's the unlock. Now let's build it.

The DBS framework, and the one piece everyone gets wrong

Every Skill folder needs a SKILL.md file. That file has three jobs, which the Growbity team calls the DBS framework: Direction, Blueprints, Solutions.

Direction is the YAML frontmatter at the top: name and description. This is the only part Claude reads to decide whether to load your Skill into context. If the description is vague, the Skill never fires. If it's overstuffed, it triggers on irrelevant tasks and burns tokens.

Blueprints are the workflow steps and rules. What Claude does, in what order, with what constraints.

Solutions are the templates and references the Skill loads only when needed. Voice samples. Hook libraries. Banned phrases. Pricing pages.

The piece everyone gets wrong is the description. Here's the rule: write it the way you'd tell a search engine when to load this Skill, not the way you'd describe it on a landing page. "A marketing assistant" doesn't trigger. "Use this skill when the user wants to write social posts, drafts for X, LinkedIn, or threads, or repurpose content across platforms in [Your Name]'s voice" triggers correctly every time.

Skills 2.0, which Anthropic shipped in March 2026, added built-in evals. You can now test whether your description triggers the Skill on the right prompts and skips it on the wrong ones. Most people don't run the evals. The ones who do ship Skills that actually fire.

The actual SKILL.md

Here's the working file. Strip out my placeholders, replace with your own, save as ~/claude/skills/marketing-engine/SKILL.md. That's the entire installation.

---
name: marketing-engine
description: Use this skill whenever the user wants to draft, refine, or repurpose marketing content across X, LinkedIn, threads, newsletters, or landing pages in [Your Name]'s voice. Triggers on requests like "write a tweet", "draft a thread", "repurpose this for LinkedIn", "create a hook", "rewrite in my voice", or any content meant for the user's audience. Also handles brainstorming hooks, generating variations, scheduling drafts, and identifying recycle candidates from the evergreen library.
---

# Marketing Engine

## Identity
- Name: [Your Name]
- Audience: [solopreneurs, indie hackers, technical founders]
- Core thesis: [the one thing your content is always pushing]
- Current offer: [what you're selling or building toward]

## Voice rules
- Sentence rhythm: short, declarative, with one long sentence per paragraph for breath
- Banned: em dashes, "let's dive in", "in today's fast-paced world", "game-changer"
- Allowed: contractions, occasional profanity for emphasis, single-word paragraphs for impact
- POV: first-person, opinionated, receipts-first

## Channel formats

### X posts
- Standalone: 240 characters max
- Thread hooks: under 260 characters, no question marks unless rhetorical
- Threads: 8-12 tweets, line breaks inside tweets for rhythm, no closing CTA
- No hashtags except in research/announcement posts

### LinkedIn
- Hook in first two lines (above the fold)
- Single line breaks between paragraphs
- 1,200-1,800 characters total
- One specific number or named entity in the hook
- End with a question only if it's genuinely open

### Newsletter
- Subject line: under 50 characters, curiosity-driven, lowercase
- Open with a scene or stat, not "hi friends"
- Sign off: [your standard signoff]

## Workflow

### When asked to write a single post
1. Confirm the channel and the angle in one line
2. Draft three variations with different hooks
3. Flag which variation is strongest and why
4. Wait for selection before refining

### When asked to repurpose
1. Read the source piece completely
2. Identify the three transferable ideas
3. Adapt format and length to the target channel
4. Preserve the strongest specific number or example
5. Never just copy-paste with line breaks added

### When asked to brainstorm hooks
1. Generate 10 hooks across these categories: contrarian, data-led, story, framework, callout
2. Mark which would work as standalone posts vs thread openers
3. Reject any hook that could have been written by a generic marketer

## References
- Voice samples: load \`references/voice-samples.md\` for tone calibration
- Hook library: load \`references/hook-library.md\` when brainstorming
- Banned phrases: load \`references/banned-phrases.md\` before any draft
- Evergreen pieces: load \`references/evergreen.md\` when recycling candidates are needed

## Scheduling
- When the user asks to schedule, defer to the configured Twitter/X MCP server
- Never invent posting times; ask if not provided
- Default cadence: 2 posts per day on X, 1 LinkedIn per weekday, 1 newsletter per week

That's the whole file. Around 60 lines. Replace the bracketed placeholders with your actual identity and rules and you have a working Skill.

The references folder is where the magic lives. Each reference file is loaded only when its workflow step needs it, which keeps your token cost low.

Mapping the SaaS replacement, line by line

Here's how the Skill above replaces each tool.

Notion AI writes drafts inside your Notion workspace. The Skill writes drafts inside any Claude session. The output gets pasted wherever you actually publish. You lose the in-Notion convenience. You gain a voice that doesn't drift between drafts because the same SKILL.md governs every output.

Jasper Pro is brand voice plus campaign templates. The voice rules and channel formats sections of the Skill are your brand voice. The hook library reference is your campaign template library. Jasper does this through a UI. The Skill does this through a markdown file you own and version with git.

Buffer is scheduling. This is the one piece a Skill alone can't do. You need a Twitter/X MCP server to push posts. The OpenTweet MCP server is open-source on npm and runs in two minutes of setup. Free tier covers solo creator volumes. You talk to Claude. Claude calls the MCP. Posts go out. Buffer's $25-a-month value lives in a free MCP plus your existing Claude subscription.

Hypefury is engagement automation, autoplugs, and evergreen recycling. The Skill handles evergreen recycling through the references/evergreen.md file: list your top 30 posts, ask Claude weekly to suggest which to recycle and how to refresh them. Engagement automation and autoplugs are the harder replacements. If those are core to your growth, Hypefury Starter at $29 still earns its keep. For most solo creators who write threads in batches, you can drop it.

Net replacement: $158 a month becomes a markdown file plus a free MCP server. The Claude Pro subscription you're already paying $20 for handles the rest.

Skills 2.0 features almost nobody is using

When Anthropic shipped Skills 2.0 in March 2026, three features landed that most tutorials still don't mention.

Built-in evals. You can write test prompts and expected behaviors directly into your Skill, then run them to see whether your description triggers correctly and your workflow produces the right output. For a marketing Skill, this means you can lock in the voice, then catch any drift the moment it happens. Run evals once a week. They take three minutes.

A/B testing. You can ship two versions of a Skill and route requests randomly between them. For a content workflow, this is how you find out whether a tighter voice rule produces better output than a looser one without guessing.

Trigger optimization. The platform now suggests changes to your description based on which prompts loaded the Skill and which didn't. If you're writing about marketing all day and the Skill only fires half the time, the trigger optimizer tells you which keywords to add.

These three features are the difference between a Skill that's a fun weekend project and one that runs your business reliably for a year.

The honest limitations

Three places this breaks. You should know them before you cancel anything.

Token cost creep. A poorly written SKILL.md loads everything into every conversation. A 500-line Skill with five 200-line reference files burns tokens fast. The fix is the conditional reference pattern: load reference files only when their workflow step needs them. The marketing-engine Skill above does this, which is why it stays cheap to run.

The 20% editing tax. Even with a tight voice profile, roughly one in five drafts needs real editing before it ships. Stormy AI flagged this exact rate in their 2026 social media guide and it matches what I see in practice. You're not getting publish-ready output every time. You're getting first drafts that feel like you wrote them, which is still a massive upgrade over generic AI output.

Things a Skill genuinely can't do. It can't watch your competitor's feed. It can't run paid ad campaigns. It can't reply to DMs. It can't track unsubscribes from your newsletter. If those tasks live in the SaaS you're cancelling, keep that one tool and cancel the other three.

Be honest about what you actually use before you cancel anything.

The bigger pattern: skills are the new dotfiles

Here's what's happening underneath this whole shift, and almost nobody is naming it.

For the last decade, developers have shared dotfiles, those tiny config files that customize their shell, their editor, their git workflow. A good dotfiles repo on GitHub got thousands of stars because it represented years of compounded preferences distilled into something forkable.

Skills are the dotfiles of the AI era. Corey Haines' marketing skills hit 6,852 GitHub stars because they're his entire marketing playbook in forkable form. wondelai's strategy skills hit hundreds of stars on the same logic. The Zovo founder is selling lifetime access to his CLAUDE.md templates for $99 because his configurations are worth that to other solo developers.

What you're building when you build a custom Skill isn't a tool. It's a config file for your own brain that an AI can read and execute. The leverage compounds the same way dotfiles compound: every refinement makes every future task faster.

The SaaS replacement is the visible win. The invisible win is that you now own a portable, version-controlled, forkable configuration of how your business operates. No vendor can take it away. No pricing page update can break it. No "we've updated our terms" email can lock you out.

That's why this matters more than the $158 a month.

What to build first, and what signal tells you it's working

Don't try to replace your whole stack in one weekend. Build one Skill, test it for two weeks, then expand.

Start with the workflow you do most often. For most solo creators, that's drafting and repurposing content. The marketing-engine Skill above is the right starting point. Fork it, replace the placeholders, run it for ten days.

The signal that tells you it's working: you stop opening the SaaS tool. Not because you decided to. Because Claude got there first and the output was good enough to ship. When that happens consistently for two weeks, you can cancel.

The signal that tells you something's wrong: you're editing every draft heavily, or the Skill keeps producing content that sounds generic. That means your voice rules are too loose or your reference files are too thin. Fix the Skill, not the tool.

Track the one number that matters: how often you reach for the SaaS in week three vs week one. If it's dropped to zero, cancel. If it's still half, the Skill needs another iteration.

Most marketing SaaS isn't expensive because it's powerful. It's expensive because you haven't yet written down how you actually want your content to sound and operate. Once you write that down, in a markdown file Claude can read, the SaaS becomes optional.

The hottest programming language for solopreneurs in 2026 is plain English. The most valuable file on your laptop is the one that tells an AI how you think.

Build that file. Cancel the rest.

Written by Guri Singh. Follow @heygurisingh for more on building a one-person marketing engine with Claude.

16 KiB Raw Permalink Blame History