AI Driven Development

Demo App

Modernization run during the presentation.

• Path: C:\Users\woute\Dropbox\Personal\Programming\UnixCode\projects\VierOpEenRij
• Repository: Laoujin/VierOpEenRij

Presentation

cd presentation
bun install
bun run dev

Update the theme:

cd presentation/theme
git pull

Actionable Items

The three

1. Install Superpowers — clone obra/superpowers into ~/.claude/skills/. You get brainstorming, TDD, debugging, code review and plan-writing skills today, no investment required.
2. Write one skill — pick the thing the agent gets wrong every week. Codify it. That's your first compound-interest payment. Skill anatomy: ~/.claude/skills/<name>/SKILL.md with frontmatter name, description. Description is what triggers loading — make it specific.
3. Wire up an LSP MCP — install Serena (⭐ 24k). Your agent gets go-to-def, find-refs, type info — not grep-and-guess. Refactors land aware of every call site. An MCP that earns its rent.

Five-minute wins

• Audit your context: run /context and /mcp, unload anything you don't need. Set up /statusline so context % stays visible. Plan to /clear at 40%.
• Move "no Co-Authored-By" to ~/.claude/settings.json → attribution.commit: "" and attribution.pr: "".
• Replace heavyweight MCPs with their CLI: gh instead of GitHub MCP, psql instead of Postgres MCP. A CLI invocation is zero schema cost; an MCP server adds tokens to every prompt whether you use it or not.
• Default to /clear over /compact. /clear plus a handover prompt keeps the next session sharp.

Fifteen-minute wins

• Add a PreToolUse hook that blocks the mistake your agent actually keeps making — git push --force to main, edits to .env, dropping migrations. Hooks run deterministically; the model can't talk its way out.
• Add a PostToolUse hook that runs tsc --noEmit and eslint --quiet after every edit. Tool-output shaping: print ✓ on success, full output only on failure. The agent gets the squiggles you get.
• Trim ~/.claude/CLAUDE.md. Delete anything inferable from the code. Big CLAUDE.md files are harmful — middle-rot eats your most-load-bearing instructions first.
• Add a top-level MEMORY.md index pointing at topic files (debugging.md, patterns.md, …). Only the index is loaded into every context; topic files load on demand.
• Set up code-review fan-out: a separate Claude session reviews diffs from your main session. The model that wrote the code is too invested to review it honestly — same context = shared blind spots.

Going deeper

• Block's 3 principles for agent skills — descriptive name, scoped trigger, layered detail.
• Latent Space — Reviews are dead — why one-shot review is a rubber stamp and how multi-agent fan-out catches more.
• Martin Fowler — Harness Engineering for Coding Agents — guides (feed-forward) vs sensors (feedback); the CAR framework.
• Agent backpressure — LSP, types, tests, lint, screenshots; controlling the flood.
• Search Kieran Klaassen on compounding engineering — the lever the talk is built around.