Planning with Claude Code

Where we are

You've picked a task, defined what "done" looks like, and built your context library. Now it's time to start talking to Claude Code. But not to produce output yet. You're still planning.

This part is about building the context.md file, the instruction set that tells Claude how to use everything you've prepared. By the end, you'll have a persona, a tool map, and a step-by-step process, all stored in one document that survives across sessions.

The context.md file

The first thing to do with a populated context library is point Claude at it and ask it to create a context.md file. Your prompt should include three things: where to look (the folder system), what you're trying to achieve (the goal and definition of good), and a design direction (recognisable references like "McKinsey clarity meets Apple design").

Claude will explore the folders, pull out brand colours, typography, and key information, and write a structured context.md in your instructions folder. This takes about 30 seconds.

Managing your context window

One of the most common frustrations with Claude Code is running out of daily tokens. Most people hit their limit not because they're doing too much, but because they're wasting context without realising it. After creating the context.md, I cleared the conversation. Here's why.

Claude Code has a short-term memory called the context window. It comes preloaded with system instructions and configurations. Every prompt you send, every response Claude gives, every file it reads, fills it up. The part that catches people out: every new prompt carries the entire context window with it. So the fuller your window gets, the more tokens each interaction costs.

Two commands help you manage this:

• /clear wipes the context window entirely. Use it at milestones, once you've saved something durable like your context.md.
• /compact summarises the window, freeing up space. Use it mid-workflow when you want to keep some continuity without starting fresh.

Defining a persona

With a fresh context window, the next question to explore in plan mode is: "What persona should my agent adopt?"

A good persona shapes both how Claude organises information and how it presents it. The trick is to combine two recognisable archetypes: one for structure, one for design. In the demo, I landed on "the ruthless clarity of a McKinsey engagement manager meets the visual precision of a senior Apple design lead."

Keep the persona clear but not over-detailed. It sets the tone, not the specifics. You want Claude to know how to think and present, not every detail of the task. Those belong in your instructions.

What doesn't work

These are common persona mistakes that either waste context or produce generic output:

• "Be professional" — meaningless. Every AI defaults to this. It tells Claude nothing about how to be professional.
• 500 words of personality traits — clutters your context window without adding value. A persona should be a sentence, not a brief.
• "Be an impact investing analyst" — too narrow. This locks the AI into one domain and limits how it can help you across different tasks.

Mapping your tools

Next, I asked Claude: "What tools are available to enhance my workflow?" The answer falls into two categories.

The best practice: map everything, install only what you need. I had Claude create a separate tool-map.md with the full catalogue, then added only the relevant tools to context.md. Nothing gets installed until it's actually needed for a specific output.

Designing the process sequence

This is where you describe, in your own words, the step-by-step process your workflow should follow. It doesn't need to be polished. I dictated mine using voice-to-text and it came out rough, but Claude is excellent at turning messy thinking into structured plans.

The key is to give Claude your raw mental model and then ask: "Is this the right way of thinking about it, or would you add other bits to make it airtight?"

Claude agreed with the core sequence but suggested an addition I hadn't considered: a storyboard checkpoint to verify narrative flow before moving into design. The screenshot verification step was my instruction, not Claude's. Here is the final sequence we landed on:

Step 4 is Claude's addition. The storyboard check catches structural problems before you invest time in design. The double screenshot verification in step 7 was an instruction I gave Claude, because it won't self-check its own output unless you explicitly tell it to.

Why HTML is the secret weapon

Large language models are trained on vast amounts of code and text. They are dramatically better at generating HTML than PDF or PowerPoint. PDF and PowerPoint are pixel-based coordinate systems designed for human users working in visual editors. HTML is a code description of layout, which is exactly what LLMs are good at.

HTML has other advantages. The same file can be styled as an A4 document, a square carousel slide, or a horizontal presentation, just by changing dimensions. It's easy to edit, supports animations, and one CSS change updates the entire document.

Choosing the right model

Claude Code gives you access to different models, each with clear trade-offs between capability and token cost.

On a £18/month subscription, you can't afford to use the most capable model for every prompt. The smart play is matching model capability to task complexity: heavy models for planning, lighter models for execution where the instructions do the heavy lifting.

What's next

You now have a context.md file with your brand system, agent persona, tool map, and a step-by-step process with quality checkpoints. You haven't produced a single output yet, and that's by design.

In Part 4, you'll hit go. You'll see the first draft, learn how to give structured feedback, and iterate to a polished output. Then in Part 5, you'll prove the whole system is reusable by running a completely different topic through the same four building blocks.