Beyond the Prompt: Claude Code

By Arpan Patel · May 26, 2026 · 27 min read

Summary (via WebFetch)

A comprehensive guide to mastering Claude Code beyond basic prompting. Covers .claude directory structure, CLAUDE.md best practices from Boris Cherny, skills as reusable expertise units, custom subagents, plugins and the marketplace, underused commands (/goal, /compact, /rewind), MCP integration, daily workflow optimization, and tips directly from the Anthropic team.

1. Claude Code Beyond the Basics

The core philosophy — attributed to Boris Cherny and the Anthropic team — is to "give Claude a way to verify its own work." Without that verification loop, the user is the only feedback signal; with it, Claude iterates until code actually runs. This is pegged as delivering a 2-3x quality improvement.

Key patterns:

• Explore, then plan, then code: Hit Shift+Tab twice for read-only plan mode. Skip planning for small fixes; use it when changes span multiple files.
• Treat plan mode like a design doc: Have one Claude write a plan, then a second Claude in a fresh session reviews it as a staff engineer.
• Reference, don't describe: Use @src/auth/login.py instead of describing files. Pipe errors with cat error.log | claude.
• Delegate, don't pair-program: Cat Wu (Claude Code team) states: "The model performs best if you treat it like an engineer you're delegating to, not a pair programmer."

Tips: Press Ctrl+G to open Claude's plan in your editor for tweaking. End prompts with "Update CLAUDE.md so you don't repeat this" after mistakes — Boris calls Claude "eerily good at writing rules for itself."

2. The .claude Directory, Properly Understood

Two scopes: Project scope (.claude/ inside the repo, committed) and Global scope (~/.claude/, rides along on every project).

Mental model: Project files describe the project; global files describe you.

Easy misses: CLAUDE.md files cascade in monorepos. Rules are path-gated via globs. Skills beat commands because skills carry supporting files, tool restrictions, and agent overrides.

3. CLAUDE.md, The Way Boris Writes It

Two principles from Boris:

• Keep it short. For every line, ask: "Would removing this cause Claude to make a mistake?" If no, cut it.
• Let Claude write rules for itself. After mistakes, tell it to update CLAUDE.md. Over weeks, this becomes a curated list of project gotchas.

3.1 The Real CLAUDE.md From the Claude Code Team

# Development Workflow
**Always use `bun`, not `npm`.**
# 1. Make changes
# 2. Typecheck (fast)
bun run typecheck
# 3. Run tests
bun run test -- -t "test name"
bun run test:file -- "glob"
# 4. Lint before committing
bun run lint:file -- "file1.ts"
bun run lint
# 5. Before creating PR
bun run lint:claude && bun run test

Boris uses Claude in PR comments to have Claude commit rules — calling this "Compounding Engineering."

Fleshed-out template: - Code style (ES modules, not CommonJS) - Workflow (bun, not npm; typecheck before done; never push to main) - Architecture (auth middleware, db queries location) - Gotchas (type distinctions, locale assumptions)

What to skip: standard language conventions, file-by-file codebase descriptions, long tutorials, API docs, anything that changes frequently.

4. CLAUDE.local.md as a Daily Driver

Sits next to CLAUDE.md, loads the same way, goes into .gitignore. After every PR, paste reviewer comments here. Two sections: project-specific feedback and personal habits. Prune after a few weeks.

5. Skills, In Depth

Skills transform Claude from "agent that can do anything" to "agent that does the three specific things your project needs, done your team's way."

A folder dropped into ~/.claude/skills/ with a SKILL.md containing frontmatter and instructions. The folder name becomes the slash command.

Three properties: 1. Progressive disclosure — only ~100 tokens of frontmatter loaded at session start; full content loads when invoked 2. Each skill lives as its own folder — can include templates, reference docs, scripts 3. Inline shell — ! runs a command at invocation time and splices output into the prompt

Frontmatter knobs: name, description, disable-model-invocation, allowed-tools, agent.

Popular skills: - mattpocock/skills (~100k stars): /grill-me, /tdd, /diagnose - Jeffallan/claude-skills: 66 language-specific profiles - Anthropic's official: /code-review, /simplify, /batch, /webapp-testing

6. Building Custom Subagents

Drop a markdown file into .claude/agents/ (project) or ~/.claude/agents/ (global). Frontmatter declares name, description, tools, and model.

Example /pr-review agent: Run git diff main...HEAD, read full files, cross-check against CLAUDE.md and rules. Flag correctness bugs, security, missing tests, N+1 queries, convention violations. Output grouped by severity with SHIP/FIX FIRST/REWORK verdict.

Popular subagents: security-reviewer, test-writer, debugger, performance-auditor, migration-writer, release-notes-writer.

Collections: VoltAgent/awesome-claude-code-subagents (100+ agents), hesreallyhim/a-list-of-claude-code-agents.

7. Plugins and the Marketplace

Plugins bundle skills, hooks, subagents, and MCP servers. Run /plugin for the marketplace browser. 1,000+ plugins, 75+ marketplaces as of mid-2026.

Day-one installs: /code-review, /feature-dev, language server plugin, /security-guidance.

8. Underused Claude Code Commands

8.1 /goal, the Ralph Loop Built In

Sets a completion condition; Claude keeps working until it holds true.

Example: /goal all tests in test/auth pass and the lint step is clean

Real examples from the article include integration tests passing without flaking 3 runs in a row, OpenAPI spec validation, docker compose health checks, and coverage thresholds.

Tip: "Combine /goal + auto mode + /focus. Write a crisp brief, set the goal, walk away."

9. MCPs as Power Tools

MCP (Model Context Protocol) turns Claude Code into a system-aware agent.

Go-to MCPs: GitHub, Context7, Sentry, Linear, Playwright, Figma, Postgres/Supabase, Slack.

Connection: Local servers use stdio; vendor-hosted use HTTP with OAuth.

Obsidian workflow: Install obsidian-claude-code-mcp, drop a CLAUDE.md at the vault root. Three tiers of memory: hot (daily session logs), warm (project notes + recent logs), cold (decisions and atoms via wikilinks).

Tip: Resist installing every MCP — bloated tool lists hurt decision quality.

10. Optimizing Your Daily Workflow

• Morning: Skim overnight output. Run /insights weekly.
• New feature: Plan mode → Ctrl+G → implement → pr-review or fresh session review.
• Bug: Reproduce first. Pipe error. Failing test before fix.
• Migrations: Use /batch — fans out to parallel agents in worktrees, each opens its own PR.
• Unfamiliar code: Hand to subagent. Keep main session clean.
• Parallel sessions: 3-5 git worktrees, each running its own Claude session.
• Writer/Reviewer pattern: Session A implements, Session B reviews fresh. Repeat until Session B stops complaining.
• Compact at milestones: /compact Preserve the decisions made, files changed, and test commands.

Tip: "Never let Claude claim success without evidence, whether that's tests, screenshots, or real command output."

11. Tips From the Anthropic Team

• "Give Claude a way to verify its output."
• Use Opus with high or xhigh effort for almost everything
• Run 3-5 sessions in parallel (worktrees beat checkouts)
• Keep a notes directory per project, updated after every PR
• Build a /techdebt slash command, run at end of every session
• Team CLAUDE.md is shared and edited multiple times a week
• Esc twice opens rewind; pair with checkpoints
• Set up Playwright MCP for UI changes
• Install a language server plugin
• Use /voice for prompting (speak 3x faster than typing)
• Auto mode + /focus + /goal — write brief, set goal, walk away
• Use Ctrl+G to edit Claude's plan before implementation
• Ask Claude to draw ASCII diagrams of new protocols and codebases

12. Resources

Official docs: Claude Code documentation, .claude directory docs, best practices, memory docs, skills, subagents, plugins, MCP, hooks

Boris and the team: howborisusesclaudecode.com, Anthropic blog on Opus 4.7 best practices, shanraisshan/claude-code-best-practice

Skills: mattpocock/skills, Jeffallan/claude-skills, addyosmani/web-quality-skills, Anthropic skills cookbook

Subagents: VoltAgent/awesome-claude-code-subagents, hesreallyhim/a-list-of-claude-code-agents

Plugins: Chat2AnyLLM/awesome-claude-plugins, claudemarketplaces.com

MCPs: Obsidian Claude Code MCP plugin, official MCP servers list, claude.com/partners/mcp

Closing Notes

The author's core insight: the mental model flipped from "I need to write this code" to "I need to set Claude up to write this code well." Setup is the work; execution is verification.

Key takeaways: - CLAUDE.md is compounding infrastructure — every mistake becomes a rule - CLAUDE.local.md captures PR feedback as free training data - Skills are the unit of reusable expertise (if you've prompted something twice, write a skill) - Subagents over kitchen-sink prompts — keep contexts clean - Parallel sessions are the most underestimated unlock