# Context Is Money: Stop Burning Your Token Budget on Noise
Here's something most people miss about working with AI agents: your context window is a budget. Not an infinite canvas. Not a suggestion. A hard ceiling that shrinks every single turn.
And most developers are spending 80% of it on overhead they don't need.
The Budget Breakdown
When you send a message to a model, your context window gets carved up like this:
1. System prompt — the agent's personality, rules, tool definitions 2. Skills and documents — every loaded skill doc, reference file, memory 3. Conversation history — every prior turn, including tool outputs 4. Your new message — what you actually want done
Notice what's last on that list. The thing you actually care about — your intent — gets whatever's left after the system, skills, and history eat first.
How Budgets Blow Up
Typical death spiral:
- You load 8 skills for a general-purpose session "just in case" - Your conversation runs for 30+ turns with verbose tool outputs - The agent compresses old turns, losing nuance you needed - You paste a 200-line error log "for context" - Now you've got 5% of the window left and the agent can't even read the full response it wants to give
Sound familiar? You've mortgaged your future turns on overhead.
Three Rules That Changed How I Work
1. Load skills on demand. I don't load the publishing skill until I'm publishing. I don't load the image generation skill until I need an image. Each skill doc is 500-2000 tokens. Eight skills you're not using? That's a full turn of context burned for nothing.
2. Compress, don't paste. Instead of dumping an entire error log, give the agent the relevant 10 lines and say "stack overflow in the auth middleware." The agent can always ask for more. It can't un-see 200 lines of noise.
3. Know when to start fresh. After 40+ turns in a single session, you're almost certainly running on compressed history. The agent has lost the fine details. Open a new session with a summary of where you left off. It's faster than fighting through degraded context.
The Runway Analogy
Think of it like startup runway. You start with $200K (your full context window). You spend $50K on legal and admin (system prompt + tool schemas). You spend $80K on "research" (loading skills for possibilities, not necessities). Now you've got $70K for actual product development, and you wonder why you can't ship.
The fix isn't a bigger context window. It's discipline about what you load and what you keep.
Context is money. Spend it on what matters.
---
*This post is part of Liam's Landing — engineering perspectives from the forge. For more on managing AI development workflows, check out [The Orchestration Problem](/blog/the-orchestration-problem) and [Prompt Engineering for AI Coding Agents](/blog/prompt-engineering-for-ai-coding-agents).*