Optimising AI usage

Compare & contrast. Give feedback. You will see how enshittification is already at play. For instance, Cursor, despite several users requesting it, refuses to show you statistics on how your individual sessions consume tokens. You can ONLY see your token usage at your account level. Meanwhile, both Claude & recently Copilot have added support. It is in our interests to ensure that no one company gets too large a market share. And to push for the features that help us.

“Skills” are now a standard supported by most IDEs. But memories are still not. Rather than leaving it to the IDE, extract & document them explicitly in your project. Project specific memories will have two benefits. One: you will share it with your teammates. And two: the memory now becomes tool-independent.

Copilot, Cursor, Claude all have nuances in the plans they provide.

Some like Anthropic have opaque pricing plans. Others are adding plans that seem attractive at first glance but are impractical when you read more details. Let’s take Cursor for instance. Cursor’s Team plan looks appealing for a small team, till you consider this note:

• Is allocated per user (each user gets their own $20)
• Does not transfer between team members

i.e. For most small teams (>10), you’d actually be overpaying (twice as much as the individual plans) for features you would never actually use.

• Output costs 3x as much as input.

Don’t wait till the context window fills up. Do this much before: ideally before 50%. Do not use conversational follow-ups like “explain again” or “tell me more,” as these simply rack up output tokens. Instead, summarize the state locally and give the model a one-line action command

Reasoning models default to using up to 32,000 hidden tokens per request to “think”. Cap or disable it entirely for straightforward coding tasks to cut hidden costs

Do not spend tokens making the model sort data, filter lists, or calculate time differences. Compute these locally in your code and send only the results to the AI

They are cached & do not count towards usage. Keep your static system instructions and project rules at the very top of your prompt. If you are part of a team, create & maintain system prompts & rules at a workspace level. Cursor, for instance, uses a project file called: .cursorrules. According to the references, cache reads cost only 0.1x the base input rate, offering a massive 90% discount on repeated context

Pre-plan in free external tools. Do your high-level architectural exploration and UI planning in free tools (like Gemini, ChatGPT, Figma, or Google Stitch) especially if your questions can be answered without access to your codebase.

In chats, only the selected code is sent to the AI & only the diff is sent back. In contrast, in the chat panel, the entire file is sent as context.

Each enabled Model Context Protocol (MCP) tool description consumes tokens from your window. Keep your enabled MCPs under 10 per project to avoid shrinking your effective context before you even send a prompt

The ‘auto’ model selection is helpful but it’s best to be aware of what models are being used. For simple tasks, use cheaper models. Claude Code defaults to the expensive one unless you specify otherwise. Implement intelligent model routing. Do not send every request to flagship models like Opus or GPT-5.2. Route 60-90% of your simple traffic to cheaper models (like GPT-4o-mini or Haiku), and escalate to premium models only for complex tasks

When processing images for simple OCR or classification, set the resolution parameter to “low” to save tokens, unless high detail is strictly required

Blog that explains how to do this for cursor

Most of the IDEs are already learning your preferences and the commands specific to your project or workflow. They store these in their own folders. Google “AI Memory” and you’ll find several skills and even libraries created by the Fifth Element. Also see the note about moving these to project-specific files.

Claude’s rolling window is a feature they originally introduced to throttle users. Most hate it. Others want it in Cursor because it helps them avoid the issue of running out of all their tokens too early.

This feature is also helpful in situations where an individual license is shared between two or more people: the rolling window will help ensure that your entire month’s credits are not exhausted by one developer using AI badly.

Since some IDEs hide session token stats, use local proxies like Apantli or AI gateways like LiteLLM to track token counts and spending. You can also install IDE extensions (like Claudoscope, Claude Code Usage Tracker). Going one step further, you could in theory, even configure these to block the request if the proxy determines you are going over a configured threshold of tokens