Stop Burning Tokens

If you find yourself running out of tokens, below are some strategies you can use to maximise your productivity.

Switch tools and make the most of free tiers

I find myself using one tool for "work" questions, one tool for personal questions and another as a coding agent.

• OpenAI ChatGPT — the free tier is perfectly good for quick web questions, brainstorming, and "explain this to me" conversations
• Claude Code — excellent for "high-level" coding (making changes with natural language)
• GitHub Copilot — great for "low-level" coding (auto-complete in the IDE, highlighting sections to provide as context)
• Gemini in the Google Suite — if you already live in Google's ecosystem.

Make the most of free tiers. Try to avoid maxing out a single tool. Rotating tools also helps to keep "memories" separate (if you use that functionality).

Use lightweight tools

Most projects don't require the most powerful cutting-edge tool. The latest models have an incredibly large cost associated with them! Lots of work can be accomplished with smaller models like Haiku and Sonnet. Make sure you really justify using the more powerful tools as they can easily churn through your allotted allowance!

Avoid Max Effort

Most LLMs let you set a "thinking" or "effort" level. High effort means more tokens and more cost. From my experience, undesirable output from an LLM is usually not an "effort" problem but an input problem.

Before cranking the effort level up, ask yourself: - Is my prompt clear? - Is the context complete? - Is the task too large? Can the task be broken into smaller chunks?

Throwing more thinking budget at a poorly framed question just produces expensive bad output.

Fix the input first. Medium effort on a good prompt will almost always outperform max effort on a vague one. Reserve maximum effort for genuinely hard reasoning tasks.

Give AI Small, Focused Chunks

"Rewrite my entire app" is a nightmare for the AI, a nightmare for your token budget, and your nightmare when the output is wrong.

Break work into small tasks: - One function at a time - One file at a time - One clearly defined problem at a time

Smaller chunks let you catch bad output early. Catching a mistake early is cheap. Catching it after five more steps of work built on top of it is very expensive.

Your Context Window is Expensive Real Estate

Avoid wasting tokens by staying out of the "dumb zone" (exceeding ~60% of your context window). You can easily waste tokens, waste time and get undesirable outputs if your context window is bloated.

Try to compact your conversation regularly. Keeping tasks small also helps with minimising context usage.

Plan Before You Let Rip

Throwing a big, vague task at an LLM and watching it fly/flail is expensive. Bad output means re-runs, corrections and wasted tokens at every step.

Take time to plan first. Write out what you want, which files are involved and what a good outcome looks like. Ask the LLM to critique your plan. You can then let the AI execute against a clear spec.

I love using tools that help to build a spec/plan and "/grillme" is a great way to make sure both you and the AI understand the task before spending lots of tokens.

Cache Your System Prompts

If you're building applications on top of AI APIs repeating the same long system prompt with every request is wasteful. Both Anthropic and OpenAI support prompt caching. When the same system prompt hits the cache the cost can drop by 90% or more.

Speak like a caveman/pirate

Fewer words. Outcome same. Caveman works.

Asking your agent to speak like a caveman is a great way of saving tokens without any degradation of output, especially when agents/sub-agents are talking to one another.

Ask for Structured Output

Output tokens are often more expensive than input tokens.

Tell the AI to answer in a structured, direct format. Asking for no preamble, no summary, and no "Great question! Let me explain..." padding means the AI gets straight to the point. You can also ask for JSON output to provide the ultimate level of structure (which is great for deterministic outputs!).

Summary

Token efficiency is about being intentional. Plan your work, use the right tool, spend the right amount of effort and reap the rewards! Spend your tokens wisely!