These aren’t hot takes from people who tried AI coding once. They’re from experienced developers who’ve spent serious hours figuring out what works—and they’ve all moved past autocomplete tools like Copilot or Cursor to agentic setups that can actually run code and iterate.
Boris Cherny: How the Claude Code creator uses Claude Code
Read on ThreadReaderApp · Reddit discussion
Cherny built Claude Code. His setup:
- Runs 5 Claude instances in terminal and 5-10 on claude.ai/code at once
- Uses Opus 4.5 with thinking for everything—slower, but needs less steering
- Keeps a shared
CLAUDE.mdin git documenting mistakes to avoid - Starts every session in Plan mode
- Has Claude verify its own work via a Chrome extension, which he says doubles or triples quality
Peter Steinberger: Shipping at Inference Speed
Steinberger’s daily workflow, distilled:
- GPT-5.2 Codex beats Claude Opus for larger tasks—it reads files for 10-15 minutes before writing, which means fewer errors to fix
- He commits directly to main
- He starts every project with CLI tools before UI
- Screenshots work better than long prompts
The takeaway: success depends less on prompting tricks and more on understanding how each model works. Steinberger also runs Claude Code Anonymous meetups—past events in Berlin, Vienna, and SF. Check his Luma for upcoming ones.
Jesse Vincent: Superpowers 4.0
Vincent builds Superpowers, a Claude Code plugin that adds structured workflows. Version 4.0 refines how Claude uses skills:
- Code review now separates spec validation from quality checks
- Skill descriptions focus only on when to use them—Claude was overconfident when given explanations of what they do
- Testing guidance is baked into the test-driven-development skill instead of scattered across multiple files
The plugin is free. Vincent enabled GitHub Sponsorships for those who find it useful professionally. It is seriously worth a sponsor.
Superpowers is my favorite Claude Code plugin so far. People keep recommending BMAD-METHOD (intro video), but it’s overkill for my needs. Other options worth knowing about: GitHub Spec Kit and OpenSpec. I still prefer Superpowers—it hits the right balance of structure without ceremony.
Simon Willison: 2025, The Year in LLMs
Willison’s annual review covers 27 trends. The highlights:
- Claude Code hit $1B run-rate revenue
- Models won gold at the International Math Olympiad with novel problems
- Chinese labs (DeepSeek, Qwen) started outpacing Western competitors on open weights
- Willison built 110 HTML+JavaScript tools using AI this year
He also discovered that asking models to draw a pelican riding a bicycle in SVG correlates with model quality.
Steve Yegge on AI coding agents
Yegge wrote the famous Google platform rant. He’s now running 20-30 AI coding agents at once through a custom orchestrator. His latest piece, Welcome to Gas Town, lays out where he thinks this is all heading.
Mitchell Hashimoto: Vibing a Non-Trivial Feature
Hashimoto co-founded HashiCorp (Terraform, Vagrant, Vault) and now builds Ghostty, a terminal emulator. He used AI to build an update notification system for Ghostty. His process across 16 sessions ($15.98 in tokens, 8 hours of work):
- Plan with an AI “oracle” before writing any code
- Break work into small, reviewable chunks rather than building complete features at once
- Create skeleton files with function names and TODOs so AI can fill in the blanks
- Know when to pivot—when AI struggled with a constraint, he switched approaches instead of forcing it
- Clean up and refactor everything AI generates
His main point: “good AI drivers are experts in their domains and utilize AI as an assistant, not a replacement.”
The pattern: developers getting the most from AI coding have rebuilt their workflows around how these tools actually work.
Update (2025-01-08): If you haven’t started with AI coding tools yet, Steve Klabnik’s Getting Started with Claude for Software Development is the place to begin. Klabnik was skeptical of AI coding until recently—he hated “spicy autocomplete” but came around once agentic dev emerged in late 2024. His advice: start with read-only analysis (paste code, ask questions) before generating anything, treat Claude as a colleague rather than a tool, and keep default permissions until you’re comfortable. Once you’ve got the basics down, the workflows above show where to go next.