Think about the context window like the session’s fuel tank. Every message you send, every agent response, every file read, and every tool loaded takes up tokens — pieces of text the model has to “keep in mind” at the same time. The tank has a fixed capacity. As it fills, quality doesn’t drop all at once: it gradually degrades. Around ~60% of use, errors start to appear. The good news: you have two pedals— /compact e /clear — to catch your breath before the tank overflows.

Illustrative diagram — the tank fills with the conversation history; start “checking the meter” and use /compact around 70% and /clear above 85% so quality doesn't plummet.

What it is

Context rot ("context rot") is the gradual decline in response quality as a conversation gets longer single session. It’s not a bug or a “tired” model: it’s a consequence of how the attention of the model works. The more text is in the window, the more the AI has to weigh at once—and past a certain point (in practice, around ~60% of use), errors start to appear.

The mechanism is easy to picture: as the conversation grows, the window fills up with noisy history — failed attempts, files you no longer use, back-and-forth. Recent messages (which often no are the most important) start to "drown out" the core rules at the beginning. The agent starts forgetting instructions, repeating what it has already done, or mixing things up—classic symptoms of bloated context.

Why learn

Key concepts

What it is

Instead of treating window usage as "all or nothing," divide it into four tiers with different actions. Think of a traffic light: 0-50% is the safe zone—work freely; 50-70% is yellow — start thinking about compacting; 70-85% is orange — run /compact proactively; 85%+ it’s the critical zone—give it /clear and start over from scratch.

Illustrative diagram — the numbers are practical guides, not exact rules. The habit that matters is: check the meter and act before before reaching the critical range.

Why learn

Thresholds turn a vague feeling (“I think this is getting confusing”) into an objective routine. You stop reacting to disaster and start acting proactively: once you hit 70%, compacting is a reflex; once you pass 85%, starting fresh is automatic. It’s the difference between driving while watching the dashboard and driving until the car dies on the road.

Key concepts

What it is

Are the two commands for managing the tank. /clear é o full reset: deletes the entire conversation and starts from scratch with empty context—use when you're switch tasks or when quality has already degraded completely. /compact é o smart summary: the agent condenses the conversation, saves what matters (decisions, current state, rules), and compresses the rest—use in the middle of a long task, when the context is high but you still need the history.

Illustrative diagram — /clear zero; /compact shrinks while preserving decisions and state. Choose based on your intent: switch tasks, or continue the same one with less overhead.

Why learn

Using the wrong command is costly: give /clear in the middle of a long task erases state you still needed (and you have to reconstruct everything by hand); give /compact thinking that it will "summarize" a change of subject leaves junk from the old task contaminating the new one. Knowing which one to use, and when, is what separates those who maintain quality from those who struggle with the agent.

/compact

Key concepts

What it is

The simplest and most powerful hygiene rule: one task per session. You focus on one goal, finish, give /clear and only then starts the next one. Don’t reuse a bloated conversation from an old task for something unrelated — that’s the fastest path to context rot.

Illustrative diagram — a task starts fresh, gets compacted midway when it grows heavy, and ends in a /clear before starting the next one. Short session = high quality.

Why learn

Clean context means accuracy. When a session contains only what matters for the current task, the agent gets more right, hallucinates less, and spends fewer tokens because it isn’t reprocessing irrelevant history with every response. Mixing tasks in the same conversation does the opposite: it dilutes focus and wastes tokens.

Key concepts

What it is

Define "done" it’s giving the agent an explicit criterion for when to stop—a definition of done. Without this, a diligent agent keeps working "just to be safe": it searches more, redoes work, expands, and burns tokens (and context) in an endless loop. With a clear target like "exactly 75 profiles", it knows the finish line and stops there.

Why learn

A loop isn’t just slow — it fills the tank. Every extra search, every “let me check again,” piles tokens into the window and speeds up context rot. Defining “done” stops the problem at its root: the agent stops at the target, the session stays short, and you don’t waste tokens or quality. It’s the cheapest way to protect context: prevent unnecessary work before it happens.

pesquise exatamente 75 perfis de <cargo> em <setor>;
pare ao atingir 75 e me devolva a lista.

Key concepts

What it is

Even before the first message, part of the tank is already taken up by things that load at the start: o CLAUDE.md, the definitions for each server MCP, system instructions. Saving context means pulling three levers: using skills (which load on demand, only when triggered, instead of living in CLAUDE.md), turn off MCP that you don't use (each server injects tool definitions into the context), and keep the Lean CLAUDE.md (fewer than ~500 lines, only what matters for every session).

Illustrative diagram — each lever gives window space back. Combine all three (plus "one task per session") and you have a routine for context hygiene.

Why learn

The context you spend at the start is context that no left for the task. A 1,500-line CLAUDE.md and five connected MCPs can use up a large part of the tank before you type the first word—and push you to the ~60% threshold much faster. Context hygiene isn’t an advanced detail: it’s what keeps your sessions long, inexpensive, and precise.

Key concepts