🪟 Scarce context window

AI doesn’t “remember everything”—it can only look at a certain amount of text at a time. That space has a name: context window (context window). It is measured in tokens, and each thing that enters the window—your request, the conversation history, the open files—uses up a slice. The problem is that too much gets in without you noticing.

Pocock’s central example is the skills. Every installed skill has a description — a sentence that says what it’s for. The model uses this sentence to decide whether to use the skill. But for power to decide, it needs see the description all the time. In other words, the description leaks for context even when the skill is never used. In Matt’s words: "every skill leaks its description into the context window." A common beginner mistake is installing dozens of skills “because they might be useful”—and each one takes up its share of the table, whether it’s empty or not.

Pocock gives the number that becomes a rule of thumb: "100 skills = 100 descriptions in context". It’s not that each skill is expensive on its own—it’s that the set descriptions is loaded alongside it, every time, so the model knows what's available. You don't even notice ten skills. A hundred skills become a block of text the model has to reread at every step just to decide what to do.

The reason is straightforward: the description exists so the model can be able to invoke the skill alone. If it’s there to be "auto-selected," it has to be visible. Add MCP servers (each one describes its tools), plus CLAUDE.md and the conversation history, and the table fills up before you even ask for anything. The common mistake: treating “I installed it” as “it’s free.” It isn’t—you’re paying in context for something you might use once a month.

There’s a switch for this: disable model invocation. Every skill can be one of two types. When the model can choose it on its own; the description has to stay visible (it leaks). When you turn on disable model invocation: true, you say: “this skill only me invoke it.” Practical result: that skill’s description doesn't leak further into context — it disappears from the wall and goes into the drawer.

Pocock gives the example of the skill "engineering zoom out": he marks it as disable model invocation, so the model never calls it on its own—only he, the human, calls it when he wants to. That fits his philosophy: he prefers procedures (that you invoke) a abilities (that the model invokes), because he wants to be “in the driver’s seat,” not delegate the thinking. In his words: "I know my skills, I don't want to delegate my thinking." The common mistake is the opposite: making everything auto-invocable “to be clever” and, as a result, filling the window with descriptions you could have hidden.

---
name: engineering-zoom-out
description: Sair do detalhe e revisar a arquitetura do sistema.
disable-model-invocation: true   # so EU invoco — a description NAO vaza
---

# Engineering zoom out
1. Liste os modulos e suas responsabilidades.
2. Aponte acoplamentos e dependencias circulares.
3. Sugira 1 refator que reduz a complexidade do todo.

Skills aren’t the only ones to blame. The CLAUDE.md (or agents.md), the MCP servers and the plugins also appear in the window. The bloat is exactly this: too many instructions piled on top of each other. Pocock sums it up in a phrase that serves as a diagnosis: "everyone bloats their context window with too much stuff."

The reason it hurts: each token of instruction is one less token for reasoning—and, worse, contradictory or irrelevant instructions confuse the model. A CLAUDE.md with 50 rules, 45 of which don’t apply to the current task, is noise. The common mistake is to “accumulate”: every time the AI makes a mistake, you add another rule. The manual grows, the results get worse, and you never remove anything. The right direction is the opposite—cut until it hurts and keep only what proves its value.

Pocock’s most radical recipe for slimming down is two-step reset (you'll see it in full in Track 5). Step 1: delete everything — every skill, every plugin, every MCP server, your CLAUDE.md, agents.md → go back to “absolute zero” and observe the agent work in this basic mode. Its wording is direct: "everyone bloats up their context window with too much stuff." See what the AI does without nothing on top.

Step 2: slowly layered back in — and that are procedures (you invoke), installed in a way that lets you customize and experiment. Bring only what you truly miss (e.g.: the superpowers brainstorming). The logic is like cleaning out a garage: after the reset, you don’t reinstall most of the skills you “thought were essential.” It’s the honest way to find out what was bloat. The common mistake is skipping step 1 — without clearing things out first, you never feel how cluttered it was.

Wrapping up the module (and Track 1): the golden rule is "each skill costs". Treat the context window as a scarce resource that you actively manage. It’s not just about clearing it once — it’s a habit: audit what leaks in, tuck what’s rarely used away in a drawer with disable model invocation, cut the bloat from CLAUDE.md, and turn off MCP servers you don’t use. Below is the hygiene checklist—copy it and run it every week before starting a major task:

CHECKLIST — context window escasso ("cada skill cobra")
[ ] SKILLS — quantas estao instaladas? Usei cada uma no ultimo mes?
[ ] VAZAMENTO — quais descriptions estao vazando a toa? (some-as)
[ ] GAVETA — as raras estao com `disable model invocation: true`?
[ ] CLAUDE.md — toda regra ainda se aplica? Corte as que nao.
[ ] MCP/PLUGINS — algum ligado que eu nao uso? Desligue.
[ ] RADICAL — se ainda estiver inchado: delete tudo, observe, recame.
Regra: cada item na janela e um item a menos de raciocinio.