This is an update on our current workflow. It is constantly evolving but this is working for us right now. A lot of people are starting to use Claude Code and perhaps this is helpful for someone diving in.

We could write a whole manual here but we've just picked out things that are really useful to know if you're just starting your journey.

If you want to use AI assistants in a disciplined way and get good results these pointers will really help.

LLM's are very literal. The clearer you are about what you want, and don't want, the better. It might also dispel some of the magic. You can ask it to make you an app and cross your fingers but we don't recommend that. It is possible to get good results if you treat it like a powerful pair programmer that requires clear direction, context and oversight. The LLM is always the assistant never an autonomous coder you leave running while you go off to play Padel. Not yet anyway.

1. The Big Mental Shift: You’re the Architect, Not the Coder

The biggest mistake people make is thinking AI is here to replace their judgment. It isn’t. Think of the AI as powerful pair programmer sitting next to you. They can type way faster than you, and they’ve read every manual on the planet, but they don't know why you’re building this specific thing.

You remain the architect and the decision-maker.
The AI handles the heavy lifting and the execution.
Quality in = Quality out. If you give it a vague "fix this," you’ll get a vague "maybe this works" fix.

2. The Golden Rule: Always Start with a Plan

It is so tempting to just start typing prompts and watching code fly across the screen. Don’t. Five minutes of planning saves you twenty minutes of cleaning up. Before a single line of code is written, think through the problem yourself, then take time to describe the problem to the AI. Make it a conversation and ask it to help you think through the approach.

Pro Tip: Ask the AI: "Come up with a plan and confirm it with me before you touch any files." This allows you to catch a bad architectural decision before it’s baked into your project.

3. Context is Everything

AI models have a "context window". Put simply, it’s how much information they can hold in their AI brain at once. If you clutter that brain with irrelevant junk, the AI gets confused and starts making mistakes.

Create a CLAUDE.md file: Put the essentials here i.e. how to run the project, your coding conventions, and your framework choices. It’s like a "Start Here" note for your AI pair programmer.
Start fresh: If the conversation is getting long and circular, it's ok to abandon it and start a new session again in a fresh context window. Revisit your plan before starting again.

4. Write Tests in the Moment

Do write tests. And don’t wait until the end to write those tests. Ask the AI to write them for the code it just finished. Because the AI still has that context fresh in its mind, it’s incredibly good at spotting the edge cases you might have missed.

5. Version Control

I'd recommend using git workflows and get into the habit of committing changes. Get the AI assistant to commit stuff regularly, ideally after each task or feature is complete and you're happy with the results. It does make it easier to rollback and start again. Even if you're not comfortable committing to a repo like GitHub, at the very least have it checking in to a local repo on your machine.

6. The Models

The models matter. We know it's not cheap but try to use the best models available/you can afford. Make sure you understand what the models are good at (and not good at) and don't be afraid to use different models for different tasks. If you're not sure try two or more LLMs in parallel to see how they approach the same problem. I regularly switch between Claude models, Opus 4.5 (for planning, big decisions, research, refactors) and Sonnet 4.5 (for those endless frontend tweak and small changes) and currently find I get good results from Gemini 3 (particularly in Antigravity) for visual frontend tasks. If you can't be bothered and you have a Max plan just leave it on Opus 4.5.

The "Zero-Stress" Workflow

If you want to move fast without breaking things, follow this rhythm:

The Chat: Describe your goal in plain English. Ask: "What questions do you have before we start?"
The Blueprint: Ask for a high-level plan (words only, no code at this point). Review it and ask the assistant to save it somewhere so you have a record.
The Work: Let it implement the plan step-by-step. Review as you go.
The Proof: Ask for tests. Run them. If they pass, you're good. Then double check and test yourself especially if it's output in a browser – AI's don't see interfaces like we do so you are always the best QA of visual UI changes.

Your First Week: A 5-Day Challenge

If you’re feeling overwhelmed, don't try to change everything at once. Just try this:

Day 1: Create your CLAUDE.md file. Just 2-3 sentences on what your project does and how to run it.
Day 2: Try the "Plan First" rule. Don't let the AI write code until you've both agreed on the plan.
Day 3: Ask for one test for every feature. See how many bugs it catches early.
Days 4-5: Use a "scratchpad.md" file to keep notes on your decisions. Fill it with anything you want to persist across sessions. You can update these anytime. The assistant will notice. Give this to the AI when you start a new session to "remind" it where you left off. Get your assistant to save stuff to markdown files especially after planning so you can stay on track.

Remember, it's a system you can tune and it does get better over time.

* Discombobulating...