Hey friend,

Today, we’ll explore

• My 8 tips for writing software with Claude Code

• OpenAI’s O3-pro model

• A new reasoning model from European AI startup Mistral

• Apple execs talk about Apple + AI

Let's dive in.

AI News

First, here’s my pick of AI news this week:

1. OpenAI Releases O3-Pro Advanced Reasoning Model, Setting New Standards for AI Problem-Solving

OpenAI's o3-pro model, now available to ChatGPT Pro and Team users, delivers step-by-step reasoning capabilities that consistently outperform previous models in science, programming, and business applications.

• Enhanced Reasoning Engine: Analyzes problems step-by-step while leveraging all ChatGPT tools including web search, file analysis, code execution, and image generation for comprehensive solutions

• Performance & Reception: Expert reviewers consistently preferred o3-pro over previous models in every tested category, particularly for clarity, comprehensiveness, instruction-following, and accuracy

• Pricing & Availability: Available now for Pro/Team users, expanding to Enterprise/Edu users next week; API access at $20/$80 per million input/output tokens (approximately 750,000 words per million input tokens)

2. Mistral AI Unveils Their First Reasoning Model - Magistral

Mistral AI's Magistral family introduces reasoning models optimized for interpretable, traceable decision-making in sectors like law, finance, and healthcare. The models generate explicit chain-of-thought logic in multiple languages, addressing enterprise needs for auditability and compliance while maintaining competitive performance benchmarks.

• Transparent Reasoning: Models perform structured multi-step reasoning with traceable logic paths, supporting eight languages including English, French, Arabic, and Chinese for global enterprise deployment

• Model Options: Magistral Small (24B parameters) available open-source under Apache 2.0 on Hugging Face; Magistral Medium in preview via Le Chat platform, API, and partner clouds

• Performance Metrics: Magistral Medium achieves 73.6% on AIME 2024 (90% with majority voting), while Small reaches 70.7%, positioning them competitively for complex reasoning tasks in regulated environments

3. Apple Execs Defend Siri Delays, AI Plan and Apple Intelligence

Apple’s AI rollout has been rocky, from Siri delays to underwhelming Apple Intelligence features. I found this recent interview with the WSJ interesting. Check it out here.

Tutorial: 8 Tips To Help You Get Started With Claude Code.

I've been using Claude Code for months now. It’s a powerful AI agent for software engineers that runs in the terminal. It's become my default way of writing code so I wanted to share some of the things I learned along the way.

Here are 8 tips to help you getting started with Claude Code.

1. Claude.md as project memory

Every project needs a CLAUDE.md file. It's like a README, but for Claude. Run /init to create one automatically, then keep it updated with your project's quirks, conventions, and context. I add design notes, coding standards, project rules. Use # during coding to add new learnings directly to the file. Keep it focused - refactor it regularly so it doesn't become a dumping ground. You can also create subdirectory-specific CLAUDE.md files and a global one in ~/.claude/ for cross-project preferences.

2. Test-driven development

Claude is really good at TDD. Have it write tests first, watch them fail, then implement until they pass. You can use XML tags to structure your prompts (Claude loves XML): <test>, <implementation>, <rules>. This forces Claude to think through edge cases before writing any implementation.

3. Commit early, commit often

Type "commit" and Claude handles everything - git add, commit message, the works. The commit messages will probably be the best you've ever seen in your career. But the real value is creating rollback points. Working authentication? Commit. About to refactor? Commit first. When Claude gets ambitious and breaks things (it will), you can just revert and try again with better instructions.

4. Have Claude review Claude

You can run many instances of Claude Code at once. Open two instances. Have one review the other's code. Fresh context means fresh perspective - the reviewer catches things the implementer missed. I've seen this catch subtle bugs and suggest optimisations I wouldn't have thought of. Create a review prompt with your specific standards and let them iterate on each other's work.

5. Don't fight bad code

When Claude goes off track, hit Escape immediately. Don't waste time trying to fix it. Just revert and start over with clearer instructions. Code is cheap to generate now. Think of it like sketching - sometimes you crumple up the paper and start fresh. Your prompts get better each time.

6. Plan mode (Shift+Tab+Tab)

The thing that really impressed me about Claude Code was it's ability to break down and plan complex work (like a human would). Have a conversation about what you want to build, create a plan, then execute. It's like pair programming where your pair thinks at light speed. Combine with -p flag: plan your feature, then claude -p "implement that user avatar system we planned" and go get coffee. Come back to review.

7. Slash commands as automation

Create markdown files in .claude/commands/ that become reusable workflows. I have commands for updating docs, security reviews, adding tests to untested code. My favorite: /fix-issue {number} - reads the GitHub issue, creates a branch, implements the fix, opens a PR. The entire workflow, automated.

8. Write PRDs for complex features

The better your context, the better Claude's output. For big features, write a proper PRD with user stories, constraints, examples. Structure it with XML tags so Claude can reference sections. Include mockups, API docs, sample data. Then reference specific parts: "implement auth flow from section 3.2."

The tooling will keep improving, but the mindset shift is what matters.

Your Opinion Matters

What did you think of today’s email? Your feedback helps me create better content for you!

Got more feedback or want me to cover a specific topic? Reply to this email and let me know.

Owain