跪拜 Guibai
← All articles
Artificial Intelligence

Why Installing More AI Skills Makes Your Model Dumber—and the One-Sentence Skill That Fixes It

By 神奇小汤圆 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

As AI coding agents become standard tools, the bottleneck shifts from model capability to instruction design. A skill that permanently occupies context window space degrades performance on everything else; a well-designed skill with a one-line interface and deferred complexity keeps the model sharp. The deletion test for skills—if removing an instruction doesn't change behavior, it was noise—gives teams a concrete way to audit their own agent configurations.

Summary

The once-dominant Superpowers framework for Claude Code is being abandoned as models improve and rigid workflows become a bottleneck. Matt Pocock's open-source skills repository offers a lightweight alternative: small, composable instructions that install only what you need. The standout is /grill-me, a one-line skill that forces an AI agent to relentlessly interview you before writing any code, eliminating the most common failure mode—misalignment between what you want and what gets built.

The system extends into a full composable workflow: /to-spec distills conversations into specifications, /to-tickets breaks them into dependency-ordered tasks, /implement runs a TDD red-green-refactor loop, and /code-review performs dual-axis quality checks. For large, ambiguous projects, /wayfinder uses a fog-of-war metaphor to create a decision map of research tickets that progressively reveal the path forward, accepting that you cannot make every decision on day one.

Underneath all of this sits /writing-great-skills, a meta-skill that applies software design principles—deep modules, progressive disclosure, and the deletion test—to the problem of writing AI instructions. It explains why every extra skill description permanently occupies the model's limited Smart Zone of roughly 120k tokens, and why the solution is not fewer skills but better-designed ones that keep interfaces tiny and complexity hidden.

Takeaways
Every skill description that sits permanently in the context window eats into the model's Smart Zone, roughly the first 120k tokens where it reasons reliably; beyond that, instruction-following degrades.
/grill-me is a single sentence that invokes a separate model-triggered /grilling skill, keeping the user-facing interface at zero context cost while the real logic lives in one place.
A router skill solves the cognitive-load problem: when you have too many user-invoked skills to remember, one router lists them all by name and use case, so you only memorize one command.
/wayfinder replaces upfront specification with a fog-of-war decision map—parent issue plus sub-tickets for Research, Prototype, Grilling, and Task—each marked with blocking dependencies.
The deletion test for skills: delete an instruction; if the model's behavior doesn't change, the instruction was dead weight and should stay deleted.
Leading words like "tight" or "fog of war" activate concepts the model already internalized during pre-training, making behavior more predictable with fewer tokens.
Skill rot comes in four forms: Sediment (stale content), Sprawl (loss of focus), No-Op (instructions for things the model already does), and Duplication (same meaning in multiple places).
Conclusions

The uninstall-Superpowers movement signals a maturation in how developers think about AI tooling: the value is shifting from comprehensive process enforcement to lightweight, composable primitives that respect the model's attention budget.

Applying software design principles like deep modules and progressive disclosure to AI instructions is an underexplored idea with immediate practical payoff. Most teams treat prompts as prose; treating them as API surfaces with interface and implementation layers changes what gets built.

The fog-of-war metaphor for project planning is a genuinely different paradigm from spec-first development. It accepts that complex projects have irreducible uncertainty and builds a process around discovering decisions rather than pretending to make them all upfront.

The observation that LLMs have a Smart Zone of roughly 120k tokens is a concrete, falsifiable claim with direct engineering consequences. If true, it means context-window benchmarks are misleading and skill designers should treat token budget as a hard resource constraint.

Concepts & terms
Deep Module
A module with a small interface that encapsulates a large amount of behavior. The deletion test checks depth: if removing the module and spreading its logic across callers causes complexity to skyrocket, it was deep; if complexity barely changes, it was shallow.
Smart Zone
Matt Pocock's term for the roughly first 120k tokens of an LLM's context window where the model reasons reliably. Beyond this zone, instruction-following degrades, making every permanently resident skill description a direct tax on model performance.
Context Load vs Cognitive Load
A trade-off in skill design: model-invoked skills reduce what the user must remember but permanently occupy context tokens; user-invoked skills keep context clean but force the user to remember and manually invoke them. A router skill resolves the cognitive-load side of this tension.
Leading Words
Words or phrases that activate concepts the model already internalized during pre-training, such as 'tight' for a fast deterministic loop or 'fog of war' for progressive exploration. Using them reduces token count and makes model behavior more predictable.
Progressive Disclosure
A design principle where a skill's content is layered by urgency: immediately executable steps go in the main body, reference material goes later, and large external references are loaded only on demand. It applies the deep-module philosophy to token consumption.
Deletion Test (for skills)
A pruning heuristic: delete an instruction from a skill. If the model's behavior does not change, the instruction was unnecessary and should remain deleted. It is the skill-design analogue of the code-level deletion test for shallow modules.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗