Beyond Vibe Coding: Advanced AI Engineering with Claude Code—Mermaid Diagrams, Stop Hooks, Rapid Prototyping

By: John Lindquist Host: Claire Vo Source: How I AI (ChatPRD / Lenny’s Podcast Network) Type: podcast

Summary

Three advanced workflows targeting senior engineers who are past the “vibe coding” phase. Workflow 1 (Diagrammatic Context): Preload Mermaid diagrams of application architecture into Claude Code’s system prompt using claude append-system-prompt "$(cat memory/ai/diagrams/**/*.md)". Diagrams can be generated from existing codebases or created when PRs close. Text-based Mermaid format is more efficient for machines to read than visual diagrams are for humans. Higher upfront token cost but eliminates slow file searches — Claude can answer complex architecture questions instantly from context. Workflow 2 (Custom Aliases & CLIs): Shell aliases for frequent commands — model switching (alias h='claude set-model claude-3-haiku-20240307'), permission modes, context loading. Beyond aliases, builds dedicated CLI tools: demonstrated sketch tool wrapping Gemini CLI for website design mockup generation with constrained terminal UX that forces focus on essential inputs (site type, page, style). Philosophy: automate repetition away, build any idea you come up with. Workflow 3 (Automated Stop Hooks): TypeScript script using Claude Agent SDK runs automatically after every Claude task via settings.local.json hook configuration. Logic flow: check for file changes → run bun typecheck → if errors, send report back to Claude with fix instruction → if pass, trigger background agent for auto-commit. In demo: Claude creates file with syntax error → hook catches it → Claude fixes → hook re-runs → passes → auto-commits. Entire check-fail-fix-recheck-pass-commit cycle without manual intervention. Shareable across engineering teams via configuration.

Key Ideas Extracted

Mermaid diagrams as compressed AI context: Text-based architecture diagrams are more token-efficient than code files and give AI instant understanding without file searches — tradeoff is higher upfront token cost for faster, more accurate responses
System prompt injection for persistent context: claude append-system-prompt loads context once per session rather than requiring @ mentions per query — architectural context persists across the entire working session
Shell aliases codify AI preferences: Simple zsh aliases reduce multi-word commands to keystrokes — model switching, permission modes, and context loading become fluid muscle memory
Build CLIs instead of wrestling with prompts: Constrained terminal UX forces focus on essential inputs — sketch tool asks 3 questions and produces design mockups vs. open-ended prompt engineering
Stop hooks as automated quality gates: TypeScript scripts using Claude Agent SDK create self-healing loops — type errors trigger auto-fix, successful checks trigger auto-commit, zero manual intervention
Scaled leverage through shared configuration: Hook configurations in settings.local.json are shareable — entire engineering team benefits from baseline quality checks without individual setup
AI as scriptable engine, not magic box: Senior engineers get value not from better prompting but from building better systems around AI — architect your environment, don’t just chat with it
Moving from 9x to 10x: For experienced developers, the productivity gain comes from automating the orchestration layer (context, checks, commits) rather than the code generation itself

Notes

Published Jan 26, 2026 on ChatPRD blog. 7-min read.
Tools: Claude Code, Cursor, Gemini CLI, Claude Agent SDK, Claude Hooks
John Lindquist is co-founder of egghead.io (developer education) and AI DX at Vercel
Links: egghead.io, AI Dev Essentials Newsletter
The stop hooks workflow is directly relevant to our own Claude Code setup — could implement similar typecheck/commit automation
Sponsors: WorkOS, Tines
This is positioned as the “advanced” counterpart to Claire Vo’s beginner episode — targets senior engineers specifically

Raw Content

Re-scraped from ChatPRD 2026-02-15. Full article content captured in Summary and Key Ideas above.