Why Do Hundreds of Skills Slow Down GPT‑5.6 Codex?

From "Stacking Skills to Boost Capability" to "Pruning Skills to Release Agent Power"

Users of Codex have observed a counter‑intuitive phenomenon: after moving to GPT‑5.6, the many Skills they previously accumulated no longer improve efficiency; instead, task execution time grows noticeably.

A simple code‑modification task takes longer.

Tasks that used to finish in minutes may now require tens of minutes.

Token consumption spikes.

The model frequently reads Skills and checks their rules.

Tool‑call frequency rises.

Complex tasks suffer from repeated planning and execution.

Many users first wonder whether GPT‑5.6’s capabilities have regressed. The analysis shows the opposite: the enhanced Agent ability in GPT‑5.6 makes the model scrutinize Skill rules more carefully, turning a large historical Skill set into an extra context burden.

Consequently, experienced developers are now deleting Skills rather than adding more. This does not mean Skills have lost value; it means their role must be re‑positioned in the GPT‑5.6 era.

1. Why Do Many Skills Degrade Performance in GPT‑5.6?

Understanding this requires a brief history of why Skills emerged.

Skills are not a permanent best practice; they evolved as an engineering tool to compensate for model limitations.

Evolution of Skill positioning:

GPT‑5.5 – Skills act as auxiliary tools that fill gaps in model capability.

GPT‑5.6 – Skills define capability boundaries and organizational norms.

The two generations have completely different Skill requirements.

2. GPT‑5.5 Era: Skills as the Model’s "Auxiliary Wheels"

In the GPT‑5.5 era, the large model already possessed strong code‑generation ability, yet it still lacked stable long‑term planning, missed steps in complex tasks, failed to choose appropriate tools, drifted in multi‑turn dialogs, and lacked reliable self‑checking.

Example workflow for building a SaaS application (ideal):

需求分析
    ↓
系统架构设计
    ↓
数据库设计
    ↓
代码开发
    ↓
测试验证
    ↓
部署上线

Actual execution often deviates:

开发页面
    ↓
修改接口
    ↓
重新调整数据库
    ↓
发现遗漏权限
    ↓
补充测试
    ↓
再次修改代码

The problem is not that the model cannot write code; it lacks a stable software‑engineering execution process. Skills therefore became an essential supplement.

What Role Did Skills Play?

Tell the model what to do first.

Specify execution order.

Fix tool‑call patterns.

Constrain output format.

Provide checking procedures.

Developers therefore accumulated many Skills across domains (frontend, backend, data analysis, PPT creation, testing, deployment, code review, etc.). The prevailing belief was: the more complex the task, the richer the Skill set, the more stable the AI execution.

3. Core Change in GPT‑5.6: Agent Capability Moves Inside the Model

The biggest shift in GPT‑5.6 is not a simple boost in coding ability but the emergence of full Agent execution inside the model.

New capabilities include automatic task decomposition, planning, tool selection, process checking, and error correction.

Previous flow (GPT‑5.5):

用户目标
↓
Prompt
↓
Skill
↓
模型执行
↓
工具调用

Current flow (GPT‑5.6):

用户目标
↓
Agent 规划
↓
模型推理
↓
工具选择
↓
执行验证
↓
结果交付

In the old model, Skills told the model the next step. In GPT‑5.6, the model decides the next step itself.

4. Why Do More Skills Make GPT‑5.6 Slower?

The core reason is that GPT‑5.6 no longer ignores Skills; it reads and enforces them seriously.

1. Skills Increase Context Cost

Each Skill adds a set of rules, usage conditions, execution instructions, examples, and constraints. When dozens or hundreds of Skills are loaded, a large portion of the context is spent on understanding which rules apply rather than completing the current task.

更多 Skill
↓
更大上下文
↓
更多 Token
↓
更长推理时间

2. Skills Can Conflict

Long‑term accumulated Skills often contain overlapping or contradictory rules. Example:

Skill A: All APIs must use REST.

Skill B: New projects should prefer GraphQL.

Skill C: Front‑end apps should prioritize Server Action.

GPT‑5.6 attempts to resolve priority, applicability, and conflicts, leading to extra analysis, rule checking, re‑planning, and verification steps that burden simple tasks.

3. High Inference Mode Amplifies Skill Cost

When operating in a high‑inference setting, the model fully reads Skills, validates constraints, and checks execution, which further increases token usage, execution time, and tool‑call frequency.

5. Re‑positioning Skills in the GPT‑5.6 Era

Large Skill collections causing inefficiency do not imply that Skills are useless. The shift is from "helping the model complete a task" to "encapsulating organizational expertise and standards".

GPT‑5.5: Skills solved the problem of the model not knowing what to do.

GPT‑5.6: Skills solve the problem of the model knowing what to do but needing to follow defined standards.

From Process‑Oriented Skills to Capability‑Oriented Skills

Recommended Skill format moves from step‑by‑step development procedures to concise organizational standards (e.g., component naming rules, security requirements, code‑review criteria, performance metrics, release requirements).

6. GPT‑5.6 Agent Technology Stack Layered Architecture

7. Division of Labor: Prompt, AGENTS.md, Skill, MCP

Component responsibilities:

Prompt – defines the immediate user goal.

AGENTS.md – specifies project‑level rules.

Skill – captures long‑term reusable capabilities for the organization.

MCP – handles external resource access.

8. How Different Users Should Manage Skills

Individual Developers

Recommended composition: GPT‑5.6 + Prompt + AGENTS.md. Avoid maintaining a large custom Skill library.

AI Engineers

Keep 3‑5 core Skills (e.g., Code Review, Project Initialization, Release Process, Enterprise Templates, Security Checks). Archive low‑frequency Skills instead of loading them constantly.

Enterprise Teams

Do not delete all Skills. Retain those that enforce governance, security, output standards, legal compliance, and brand guidelines.

9. Future Trend: Fewer Skills, Higher Value

Instead of 100 ordinary Skills, the future will likely involve around 5 high‑value Skills that encapsulate organizational knowledge, while the model handles general intelligence and the Agent manages planning and execution.

10. Summary – From "Constraining the Model" to "Designing Agents"

In the GPT‑5.5 era, developers focused on telling the AI each step, leading to heavy investment in Prompt writing, Skill creation, and process solidification.

In the GPT‑5.6 era, the focus shifts to enabling the AI to autonomously complete tasks within defined boundaries. The core of AI engineering becomes designing an Agent system that knows what the Agent, Prompt, AGENTS.md, Skill, and MCP each own, rather than simply stacking the most Skills.