Tech • AI • Robotics • Game
Your custom AI agent is a waste of time (use skills)
7/10TL;DR
Building separate AI agents for every department is often inefficient because many of them share the same core functions, while the real differentiator lies in reusable skills that capture company-specific expertise.
KEY POINTS
One core agent behind many use cases
Companies are increasingly trying to build separate AI agents for sales, legal, marketing and other teams, each with its own prompts, tools and code. But the underlying agent often performs the same basic actions: reading files, running code and calling tools. The main variation is not the agent itself, but the operational knowledge layered on top of it.
The value shifts from agents to skills
Engineers working on Anthropic’s Claude Skills concluded that the common foundation across specialized sub-agents was largely identical. In this framing, the LLM acts like a processor, the surrounding runtime such as Claude Code or Codex acts like an operating system, and skills function like applications. What changes business outcomes is therefore less the base model architecture than the specialized capabilities attached to it.
What a skill actually is
A skill can be as simple as a folder containing one file with instructions for how to perform a task, plus optional scripts. At startup, the agent only needs to read the skill’s name and description, opening the rest only when necessary. That design reduces wasted context and allows the model to access targeted procedures on demand.
Reusable scripts improve consistency
One recurring problem with AI assistants is that they often rewrite code from scratch for repeat tasks such as reformatting a presentation or extracting information from a PDF. That can produce inconsistent outputs from one run to the next. Saving a script that produced a satisfactory result inside a skill allows the agent to reproduce the same process more reliably instead of reinventing it each time.
Descriptions must be precise
Skill discovery depends heavily on clear descriptions. If two different skills are both vaguely labeled as helping with “content creation,” the model has little basis for choosing the right one and may select arbitrarily. A skill the model cannot identify correctly is effectively unusable, making precise naming and scoping critical.
Corrections should be turned into upgrades
Users frequently correct an agent’s output when working in environments like Claude Code or Codex. Those corrections become far more valuable when they are incorporated back into the relevant skill. Otherwise, the lesson disappears at the end of the session and must be retaught repeatedly, preventing cumulative improvement.
Verification steps should be built in
Stronger skills do not just define how to act; they also define how to check the result. Recommended verification steps include opening source material, reviewing rendered outputs for a website or video, and comparing the final product with the expected result. The goal is to guide the system toward a validated fourth version rather than accepting a weak first draft.
Skills do not replace model quality
Skills can improve workflow reliability and encode expertise, but they do not transform a weaker model into a stronger one. A skill cannot give a lower-tier model the reasoning ability of a more capable system such as Opus. Nor do skills improve on their own: they require human input to encode better practices and raise performance over time.
CONCLUSION
The main strategic choice is not whether to build more custom agents, but how to capture operational know-how in reusable skills. Organizations that treat skills as durable, verifiable layers of expertise may gain more consistency and efficiency than those endlessly rebuilding the same agent in different forms.
Ask a question