AI coding assistants changed how we build software. They also introduced a new kind of waste: tokens. Every word the AI reads and writes costs money and time — and most workflows burn far more than the work requires.
After hitting usage limits one time too many, we rebuilt our workflow around four habits:
1. Compress the conversation
The assistant doesn't need pleasantries or filler to write correct code. Stripping conversational padding from AI output cut our output tokens by well over half with zero loss in accuracy.
2. Index the codebase — don't re-read it
A question that used to take the AI 40+ file searches now resolves in 2–4, because we keep a pre-built knowledge graph of the whole codebase. The AI queries the graph instead of re-reading files it has seen a hundred times.
3. Review changes through the graph too
Code review with AI used to mean re-reading entire files. Reviewing a change through the same graph — what changed, what it impacts, who calls it — keeps reviews sharp and cheap.
4. Force the lazy solution
The best code is the code you don't write. We push our AI tools to produce the simplest thing that actually works. Less code means fewer tokens, and — more importantly — fewer bugs.
The result
Same output, roughly half the cost, and faster iterations. Clients don't see the tokens — they see that we ship in days. This is why.
