Evaluating advanced Claude Code patterns for “doing more with less” — maximizing output through better prompting, parallelism, automation, and skill design.
Why Evaluating
Claude Code is the primary development and knowledge-management tool. Every efficiency gain here compounds across all projects. Currently using it for vault ops, Obsidian skills, code generation, and research synthesis. Room to push further.
Current Usage
- Obsidian vault management (skills, templates, daily logs)
- Code generation and project scaffolding
- Research synthesis and analysis
- MCP server integrations (Calendar, Gmail, Notion, Supabase, etc.)
Patterns to Explore
- Boris Cherny’s 13 tips — parallel sessions, plan mode, slash commands, hooks
- Custom slash commands for repeated workflows
- Subagent patterns (code-simplifier, verify-app equivalents)
- PostToolUse hooks for formatting/linting
- Background agents for long-running tasks
- Teleport between terminal and web sessions
Notes
Community Skills via npx — 2026-04-05
A list of installable Claude Code skills, all installed via npx:
Essential (any project):
| Skill | What it does |
|---|---|
| Superpowers | Full dev cycle: planning → subagents → TDD → code review. 40K stars |
| Frontend Design | Removes “AI slop” — forces bold, unique design |
| Vercel Web Design Guidelines | 100+ accessibility and UX rules. 133K installs/week |
| Vercel React Best Practices | 57 rules for React/Next.js — waterfalls, bundle size, re-renders |
| Webapp Testing | Claude tests your local site in a real browser |
For larger projects:
| Skill | What it does |
|---|---|
| Trail of Bits Security | Professional security audit: CodeQL, Semgrep, vulnerability scanning |
| SEO | Full SEO audit |
| Document Skills | Creates real PDF, DOCX, XLSX, PPTX files |
| Vercel Composition Patterns | Proper React component architecture |
Evaluation Results — 2026-04-08
Installed:
- Frontend Design (already had via
claude plugin install) — 400 tokens, forces bold aesthetic direction before writing any frontend code. Zero overhead. - Vercel Web Design Guidelines — 100+ accessibility/UX rules, fetches latest guidelines each time. Installed via
npx skills add vercel-labs/agent-skills --yes --global. - Vercel React Best Practices — 57 performance rules ordered by impact (waterfalls, bundle size, re-renders). Same install.
- Vercel Composition Patterns — came bundled with above. Proper component architecture.
- Also installed: React Native skills, React View Transitions, Deploy to Vercel, CLI with tokens.
Skipped (with reasoning):
- Superpowers — 22K tokens at startup (11% of context). Forces full brainstorm→TDD cycle on every task regardless of size. A custom
/implement-specskill covers the core value (read plan, execute step by step) without the overhead. - Trail of Bits Security — Requires CodeQL + Semgrep installed locally. Designed for professional security auditors. Overkill for early-stage projects. Revisit when apps handle customer data at scale.
- Webapp Testing (Playwright) — High token cost from browser output serialization. Requires Playwright. Useful for complex UIs but lower priority. Revisit when a project needs real browser testing.
Custom skills built (2026-04-08):
/implement-spec— reads a spec file, parses steps, implements sequentially with checks/verify— auto-detects project type, runs lint/types/tests/build, reports results/commit-push— detects git remote/branch, reviews changes, commits and pushes with confirmation/deploy— reads deploy config from project CLAUDE.md, SSHs to VPS, pulls and restarts
New Features Spotted — 2026-04-12
From @jens.heitmann reel (“Anthropic silently shipped 3 more features”):
/ultraplan— dispatches multiple explorer agents in parallel for research before planning. Shows “3 explorer agents dispatched” then synthesizes findings into a plan.- Power-ups — gamified onboarding (1/10 unlocked). Teaches one Claude Code capability per power-up: @ file refs, shift+tab modes, /rewind, background tasks, CLAUDE.md/memory, MCP, skills/hooks, subagents, /remote-control + /teleport, /model + /effort.
/insights— analyzes usage across sessions (“Analyzing 84 sessions from last 30 days”). Session analytics for patterns and optimization.
Claude Code Desktop — Parallel Agents Rebuild — 2026-04-20
From @evolving.ai reel (“Another major week in AI”). Claude Code desktop has been rebuilt around parallel agent sessions.
What it does:
- Multiple Claude coding sessions run side-by-side in one window with a sidebar to manage them all
- Kick off a refactor in one repo, a bug fix in another, tests in a third — switch between them as they finish
- Sidebar sections: New session, Scheduled, Customize, Pinned (e.g. “Build the alignment grid demo”), Recents (e.g. “Migrate API client to fetch with retries”, “Fix race condition in upload queue”, “Add keyboard shortcuts to command…“)
- Background Tasks panel — shows running agent work (e.g. “Verify constraint solver convergence” running 1m 12s, “npm run dev (background)” running 8m 03s)
- Inline preview pane with errors/diffs alongside the editor
- Plan pane — “Select any text to leave a comment for Claude”
- Auto-accept edits toggle
- Default model: Opus 4.6
Why this matters — HIGH: Directly maps to a multi-project workflow — research sprints, build work, and drafting that today happen in parallel via separate terminal tabs and manual context switching. This collapses that into one window with scheduled runs and background tasks. Scheduled/recurring agent work is also a demo-able feature for the “agent as coworker” concept.
part of tooling