How I teach Claude Code to see, think, and do
Imagine: you have a newborn. Smart. Capable. But can't do anything yet.
You need to give it eyes — to see. Hands — to do. And a brain — to understand what exactly to do.
I've already written about memory. CLAUDE.md, project files, context after compaction. That's the baseline. But without skills, your agent is a blind intern with a good memory.
Eyes
By default, Claude Code can't see websites. It works with code, but it doesn't see how things look in the browser. It doesn't see what the user sees.
For that, you need a headless browser. I tried several — Chrome MCP Server works best.
Once installed, the agent can open any page, look at it, take a screenshot, check how it looks on a phone.
Without it, he guesses. With it, he sees.
Hands
Design is Claude Code's weakest spot out of the box. He builds pages that work. But they look like they were coded in 2008.
Anthropic's built-in skill (frontend-design) is the baseline. It's not enough. I add two more:
interface-design — the art director. Remembers design decisions across sessions: colors, fonts, spacing. Knows how to critique a layout and explain what's wrong.
ui-ux-pro-max — the library. 67 styles, 161 palettes, 57 font pairings. When you say "make it pretty" — he doesn't guess, he picks from proven combinations.
Plus anti-lazy — that's not a skill but a hook. A rule that fires on every message: read the code first, then change it; don't half-ass it; fix things for real. Without it, the agent sometimes makes shallow edits and says "done."
Brain
So the agent can comprehensively analyze any site, I split the work into 4 layers.
-
Technical foundation — so the site can rank at all. Speed, indexing, structure, broken links, page connectivity. technical-seo-checker, on-page-seo-auditor, internal-linking-optimizer
-
Content — so Google understands what you're about. Text quality, headings, descriptions, schema markup, refreshing stale pages. content-quality-auditor, meta-tags-optimizer, schema-markup-generator, content-refresher
-
Market — so you hit actual demand. What people search for, what competitors have, what you're missing. keyword-research, competitor-analysis, content-gap-analysis
-
AI visibility — so you get cited. ChatGPT, Perplexity, Google Maps, Wikipedia connections. geo-content-optimizer, entity-optimizer
Plus separately — tracking and reports: rank-tracker, performance-reporter.
Each layer gets its own skills. Each skill is a set of checklists and instructions the agent follows on a specific slice of the task.
Why one agent isn't enough
This is the most important part.
One agent tries to hold everything in its head — tech, content, competitors, design. And at some point it starts to: — forget context — mix tasks — give shallow solutions — skip pieces but say "done"
That's not a bug. That's the limit of the context window.
So I run 3–4 agents in parallel. Sometimes on different servers. One — for tech. The second — for content. The third — for keywords and competitors. The fourth — for design and conversion.
They work simultaneously. Results merge into one report.
This isn't about "more powerful." It's about each agent holding its narrow slice and not getting distracted by the rest.
Eyes, hands, brain. Without skills, Claude Code is smart but blind and clumsy. With them, it's a working system that does in days what an agency does in months.
In future posts I'll show how this looks on real projects.