The Harness Around an LLM Delivers Multiplied Gains: What Changed in Claude Code Over a Year

After a year and a half of heavy use of Claude Code, dozens of experimental runs, and watching a team of engineers at work, one experienced developer reached an unexpected conclusion: the main lever for improving LLM quality is neither the model generation nor an upgrade to a newer version. It is the complex system built around the model.

What is a harness

In the English-speaking engineering community, this surrounding system is called a harness: the full environment in which a language model operates. It isn't just a prompt typed into a chat, and it isn't a single behavioral rule. It's a complex system that governs how the model interacts with the outside world: its rules, its boundaries, the tools it can use, and how much context it can see. Without this environment, even the most powerful model works inefficiently. Think of an excavator without controls: it has an engine, but nothing to dig or move earth with.

Why the harness delivers bigger gains than new models

A year of observation showed that switching to a newer model version brings a noticeable but limited quality gain. Moving from Claude 3 to Claude 4 is an improvement, but it amounts to a few percent of performance. Meanwhile, every new harness layer around the same model produces a multiplied jump.

Add a system prompt that clearly defines the model's role, and quality jumps. Connect tools that let the model interact with the real world, and it jumps again. Expand the context, add skills, configure permissions, and each time quality grows several-fold. That shifts engineers' focus away from chasing new versions and toward designing the environment the model runs in.

Components of an effective harness

An effective harness consists of several key layers, each contributing in its own way:

• System prompt — explicit instructions that define behavior, communication style and boundaries
• Tools — model access to external APIs, databases, browsers and compute resources
• Context — sophisticated management of what the model sees, remembers and can rely on
• Skills — pretrained patterns and algorithms for solving common tasks
• Hooks — events and triggers that fire under specific conditions
• Permissions — the line between what the model can and cannot do, and which files it may read and write
• Memory — long-term and short-term retention of project context, decisions and conclusions

Each of these layers can be optimized independently, improving performance for a specific type of task.

"The power is there, but there's nothing to use it with" — that's how

an experienced engineer describes a model without a harness.

What this means

Developers are shifting their focus from the race for newer, more powerful models to building an intelligent environment around the ones they already have. Investments in system prompt engineering, context management, skill development and permission tuning pay off faster and deliver more than waiting for the next model version. This opens up new ways to customize for specific tasks and teams without requiring costly hardware upgrades or licenses.