What triggered this for me was mundane. A fresh chat knows nothing about my projects, my priorities, my writing, my folder structure. So you explain it again, every time: ten minutes of typing context before the first useful line arrives.
The obvious fix is a mega-prompt: everything in one block, pasted at the top, fingers crossed. That scales right up to the point where you have two tasks needing different knowledge. A prompt that explains your code style, your brand voice and your invoice templates at the same time ends up explaining none of them well.
What helps is a different structure. You split your setup into departments. Each department is a folder with skills and a config, and above them sits one file defining who you are and what matters. Set it up once and the system works with you every day.
Seven departments, one CEO file. That's the whole map:
| Department | What for | Core tools |
|---|---|---|
| 01 Engineering | Plan, build, review | Superpowers, Context7, claude-mem |
| 02 Design | Look and quality control | ui-ux-pro-max, design-review |
| 03 Marketing | Copy, SEO, voice | copywriting, ai-seo, brand-voice |
| 04 Social | Find topics, ship content | FIND3X, POST3X |
| 05 Video | Images, B-roll, compositions | Higgsfield, HyperFrames, Remotion |
| 06 Office | Mail, calendar, knowledge, bank | MCP connectors |
| 07 CEO | Context and priorities | one CLAUDE.md |
Technically every department lives in the same directory. Skills go under ~/.claude/skills/, each as its own folder with a SKILL.md inside. Claude loads every skill's description and only pulls the full content when it matches the task at hand. That's why a large library costs you no context until you actually use it.
Superpowers is a skill bundle that prescribes the process: brainstorming before implementation, test-driven development, code review as its own step, systematic debugging. The effect is less spectacular than it sounds, and that's exactly why it works. Claude clarifies what should be built first, then types.
Context7 attaches as an MCP connector and serves current library documentation. That solves the problem that every model has a knowledge cutoff while framework APIs move faster than that cutoff.
claude-mem is memory across sessions. It stores observations from past runs and makes them searchable, so a new session knows what the last one was working on.
Your first step: install Superpowers and deliberately say "let's build X" once, before you ask for a specific file. You'll see straight away whether the forced planning step helps you or gets in your way.
ui-ux-pro-max ships 50 style directions, 21 colour palettes and 50 font pairings, plus patterns for React, Next.js, Vue, Svelte, SwiftUI and Tailwind. The value is that you can name a direction instead of circling it. "Swiss, lots of white space, one accent colour" is enough.
design-review runs afterwards: it checks the built interface for inconsistent spacing, broken hierarchy and the usual AI tells, then fixes them in the source.
If you want your own design system, Claude Design System is the faster route there.
Your first step: run design-review against a page you already have. The report tells you what's genuinely off about your current look.
Three skills that work together. copywriting for landing pages and sales copy. ai-seo for visibility inside AI answers, which is a different game from classic Google ranking. brand-voice keeps your tone consistent across channels by deriving a style profile from real texts of yours.
The difference from "write me some copy" is in that last line. A voice profile built from your own posts sounds like you. Three adjectives in a prompt do not.
Your first step: collect ten of your best pieces and have a voice profile built from them. That file is what every other marketing skill reads afterwards.
This is my own stack. FIND3X scans a curated channel list weekly, scores posts by relevance, performance and recency, and puts the strongest topics up for selection. POST3X takes a chosen topic and turns it into captions, carousels, reels or blog posts, then publishes them.
This article went through exactly that pipeline: produced as a carousel first, then run through the blog side of POST3X into the text you're reading.
Your first step: write down a topic you were going to post this week anyway, and build a skill that turns it into your standard formats. Two formats are plenty to start.
Higgsfield generates images, B-roll and video clips. HyperFrames turns them into compositions, meaning videos described as HTML and then rendered, captions and text-to-speech included. Remotion is the option for anyone who'd rather write their videos in React.
All three connect through their official skills, for example:
npx skills add heygen-com/hyperframes
Your first step: build one clip from a text you already have. Start with a paragraph you've already written. The campaign comes later.
This department holds no skills, only connections. Mail, calendar, knowledge base and bank account attach directly as MCP connectors. MCP is the standard Claude uses to talk to external services: instead of you copying between tabs, Claude executes.
That makes sentences possible that used to cost two tools and ten minutes: "What's on this week and which invoices are still open?"
Which connectors pay off first and how to set them up is covered in 5 MCP Services.
Your first step: connect exactly one service you touch daily. Calendar or mail. Not all of them at once.
This is where most setups fail, and the one part you can't skip.
A CLAUDE.md is a Markdown file Claude reads automatically before doing anything. It answers three questions: Who are you? What's the current priority? How do you sound?
It exists on two levels. Globally at ~/.claude/CLAUDE.md it applies everywhere: your working style, your language, your code standards. Inside a project folder it applies only there: architecture, conventions, open decisions.
A workable skeleton:
# Who I am
Role, company, what I'm working towards.
# Priorities
1. What pays right now
2. What's next
3. What's deliberately waiting
# How I work
Short sentences. One concrete proposal, not five options.
Tell me when my approach is bad.
# Stack
What I build with, and what I deliberately avoid.
Without this file the rest is a toy. Seven departments with no shared context are seven strangers using the same tool.
How to structure the file in detail is covered in Using CLAUDE.md properly.
Not all at once. The order matters more than the completeness.
CLAUDE.md. Twenty lines is enough, you'll keep extending it anyway. Everything after this reads that file.You install skills either through the package route (npx skills add <owner>/<repo>), through plugin marketplaces in Claude Code, or by hand by dropping a folder with a SKILL.md under ~/.claude/skills/. The manual route matters most, because that's exactly how your own departments come into existence.
No spam, unsubscribe at any time