Opus 4.5 is out and it feels like it's a big step forward in AI innovation.

The new model outperforms other models when working on long-context tasks and more complex reasoning chains. It's great for long coding tasks and multi-file refactors, but it's also great for creating working solutions for other parts of the business too.

Maybe you're not convinced and still on the other side of the fence when it comes to AI. Regardless of what you think of the model benchmarks, the tools are good enough now that waiting to invest in AI is the riskier bet.

Getting Started

My team has been experimenting with AI a lot recently, and we're starting to see some great results. But if you're still new to the AI scene, and confused where to start it can be difficult to know what to do.

1. Get a Subscription!

Give your team Claude Code subscriptions to get them going!

Ensure they have a safe place to start exploring how to use the tool and play with how it works. Some engineers feel threatened by AI, whilst others over-trust the output. You can acknowledge this by framing Claude Code as a "junior pair programmer that needs review".

Avoid starting with a production-critical feature and instead begin experimenting on internal tools or fixing bugs instead.

$200/month for a license for developers will quickly pay itself back when they get used to the tools.

2. Knowledge Sharing

Give developers a way to share their learnings with one another, and start to build a shared context domain with the things that work within your repo. This could be through a shared Slack channel or regular catch ups. Encourage the team to talk about AI.

3. Guardrails

All the old rules to software development still apply.

Guardrails are incredibly important. The AI will inevitably go on tangents and will still write bloated code from time to time. AI isn't an excuse to start throwing out the rules of good code and PR hygiene.

Code should follow standard patterns, have explicit naming and organisation conventions, and a clear separation of concerns. PRs should still follow best practice rules. Keep code changes small. Keep writing tests. Keep reviewing code.

Tests aren't just a safety net, they're the feedback loop that keeps the agent honest. Consider utilising TDD to get the AI to write code to pass the tests.

Creating an AI bot (or using a tool like CodeRabbit) with your Github pull requests to check the PR against the standards will give you an extra line of defence, but human validation is still incredibly important.

4. Standards

Keep documenting your coding standards.

Create a standards document within your code repository that your team and your AI works from - covering things like database patterns, coding principles, security requirements. Better documentation of your code base will lead to more predictable results from your AI agents.

Full-stack ownership starts to become more natural with AI, you may find engineers moving out of their front end or back end specialities to work more efficiently; so it's important they (and their AI) can understand how to do that with this shared foundation.

Grow this document over time from the things that frustrate you.

5. Build Agent Skills

Once your team has the basics down, you can start building reusable skills for common tasks. For example, scaffolding a new service, writing tests for existing code, generating PR descriptions, or setting up a new feature end-to-end.

These are essentially prompt templates or task definitions that live in your repo alongside your standards document. They're how the compounding gains start to show up.

Make sure you are saving shared agent context with your PRs, so instead of every developer figuring out how to prompt the agent for a database migration, someone can write it once and everyone will benefit.

Conclusion

The cost of experimenting is low, the iteration cycles are fast, and the most expensive thing you can do right now is stall. Teams that are seeing 10x productivity gains didn't plan their way there — they started small and built confidence. The best time to start with AI was several months ago, the second best time is now.