Series: A Practitioner's Guide to AI-Assisted Development

The journey from AI-sceptic to agent-native isn't a smooth ramp — it's a series of uncomfortable jumps. This post maps the stages of AI-assisted development and identifies three critical fulcrums where developers get stuck.

AI coding agents have made code production nearly free. Drawing on insights from Kent Beck, Paul Ford, and Simon Willison, this post argues that the value has shifted from writing code to defining systems — contracts, invariants, SLAs, and verification.

A rigorous study found AI-assisted developers are 19% slower yet believe they're 24% faster. The gap reveals that AI is best understood as an exoskeleton — amplifying human judgment rather than replacing it — and that the real risk isn't bad code but cognitive debt from velocity without understanding.

The setup checklist I wish I'd had on day one with Claude Code — isolation and sandboxing, terminal configuration, LSP integration, token monitoring, and what not to install.

CLAUDE.md is the single file that determines whether your AI coding agent shines or flounders. Research shows auto-generated context files hurt performance — what works is a minimal, human-curated briefing containing only what the agent cannot discover on its own.

Every Claude Code session starts from scratch. Teresa Torres's three-layer context system and Patrick Zandl's claudecode-kb offer two practical approaches to the same insight -- memory is a design problem, not a tooling problem. The key habit is stop and capture, turning every repeated explanation into a reusable context file.

Your prompt is 0.1% of what the model sees; the other 99.9% is context engineering. A four-layer framework for thinking about prompts, practical guidance on what belongs in context and what doesn't, the invisible context problem from a million-line codebase, and why agent-controlled retrieval beats RAG for cross-file reasoning.

AI agents amplify whatever engineering process they're given. Bad specs produce confident garbage at machine speed. Three practitioner workflows for separating planning from execution: the normal coding flow, Boris Tane's annotated plan cycle, and Jamon Holmgren's Night Shift spec discipline. The pattern is simple; the discipline is resisting the urge to skip straight to implementation.

The context window is a budget, not a feature. Auto-compaction hides the bill until the agent starts hallucinating. Practical tactics for staying under budget: scope per session, offload to disk, dispatch subagents for research, and clear aggressively between phases. The goal isn't a bigger window; it's needing less of it.

Load 84 MCP tools and 15,540 tokens are gone before you ask a question; after thirty minutes you've burned 40% of your context on tool definitions you didn't use. Holmes and Yilmaz make the case for CLI-first, and I've mostly come round: CLIs are debuggable, composable, and 92-98% cheaper in tokens. MCP still earns its keep for a few tools, but the default should flip.

Skills are Standard Operating Procedures the agent loads only when needed — progressive disclosure applied to AI context. Without a forced-eval hook they activate 55% of the time; with one, 100%. That gap is the difference between skills working and skills being decoration. Plus why hooks are the enforcement layer that makes any of it reliable.

A complete guide to provisioning and hardening a Hetzner VPS for AI coding agents — covering cloud-init security, Tailscale mesh networking, Cloudflare Tunnel, Claude Code resource planning (including the subagent memory pitfalls), and automation scripts to go from zero to a working environment in minutes.

A curated configuration for Claude Code that bundles plugins, skills, CLI tools, and sandbox guards into a single install script. What it does, how to use it, and why you should fork it immediately.