Small habits, big context windows

The way you use AI affects more than just your output.

In my last post, I asked whether Claude is as bad as a banana – and admitted, honestly, that we don’t fully know yet.

What we do know is this: the carbon cost of an AI prompt isn’t really about the prompt. It’s about everything the model is carrying when it answers you.

That distinction matters more than it might sound. And once you understand it, a handful of small habits start to look a lot more worthwhile.

The email problem, but for AI

There’s a useful parallel in the carbon footprint world. A single email has a tiny footprint – around 0.03 grams of CO2. A traditional letter, by contrast, is significantly higher. So emails are better for the environment, right?

Sort of. The problem is that we send billions more emails than we ever sent letters, because they’re so cheap and easy. The individual footprint is low; the aggregate impact is enormous. Convenience drives volume, and volume drives emissions.

AI has the same dynamic, but with an added twist. It’s not just that we’re prompting more often – it’s that each prompt carries more weight than it looks like. Every time you send a message in an ongoing session, the model doesn’t just process what you just said. It processes your entire conversation history – every previous message, every response, all the context it’s been given, every tool or integration it has loaded. That’s what’s generating the tokens, and tokens are what cost both money and energy.

So a single message that looks like six words might actually represent hundreds of thousands of tokens of context being processed underneath it. Which means the way you manage your sessions matters as much as what you put in your prompts.

What we changed, and why

When we started thinking about this at Hiyield a few practical habits emerged. None of them require you to use AI less. They just require you to use it a bit more deliberately.

Start a fresh chat when you switch topics

This is probably the single most impactful habit. A session that started as a brand strategy brainstorm and drifted into writing copy and then turned into a technical query is carrying all of that history with every message you send. The model doesn’t need that context for the technical query, but it’s processing it anyway. Starting a new chat costs you nothing except the habit of remembering to do it.

Be specific about the length of the answer you want

If you need a one-paragraph summary, say so. If you want bullet points rather than a full explanation, ask for them. Left to its own devices, a model will tend towards thoroughness – which is often more than you actually need, and always more tokens than a targeted answer. Specificity isn’t just good prompting practice; it’s more efficient.

Use /clear regularly if you’re working in Claude Code

For anyone using Claude in the terminal, context builds up fast – and it builds up in ways that aren’t visible. A session running all day on a single project, without ever clearing the context, can accumulate tens of millions of tokens. Most of that is memory overhead, not actual useful work. A quick /clear between tasks resets the window and keeps things lean.

Share IDs directly rather than letting the model search

This one is particularly relevant for anyone using Claude alongside project management tools. If you want Claude to pull up a specific task, brief, or piece of work, giving it the direct ID (a ClickUp task ID, for instance) is dramatically more efficient than asking it to find the thing by description. Search queries generate a lot of token overhead. A direct reference doesn’t.

Think about when you’re prompting, not just how

This is a more experimental one, but worth knowing: the carbon intensity of the UK electricity grid varies throughout the day. Using AI during lower-intensity periods – typically overnight or mid-morning rather than early evening peaks – means the electricity powering those requests is likely to come from a greener mix. It’s a small lever, but it’s a real one.

This isn’t about using AI less

I want to be clear about what this post is and isn’t arguing. It’s not a case for prompting less, or for being precious about every interaction, or for manually calculating the carbon cost of asking Claude to help you write an email.

The goal is conscious use rather than guilty use. Understanding that context windows accumulate, that session length has a real cost, and that a few small habits can meaningfully reduce that cost – without reducing the value you’re getting – feels like useful information. Not a burden.

We use AI heavily at Hiyield, and we intend to keep doing so. The question we’re trying to sit with is how to do it in a way that’s honest about its impact and genuinely committed to improving. These habits are a small part of that answer.

The bigger parts – better measurement, more accurate carbon accounting, and eventually a clearer picture of what our AI use actually costs – are still in progress. But in the meantime, /clear is free.

This is the second post in a series about AI and sustainability at Hiyield. The first – Is Claude as bad as a banana? – covers how we’re trying to measure our AI carbon footprint and why it’s harder than it should be.

If you’ve found better approaches or want to compare notes, we’d genuinely love to hear from you.