YOUR AI AGENT ISN'T DUMB. IT FORGOT HOW TO USE YOUR TOOLS.

You have felt this. Your agent nailed a task yesterday. Today it fumbles the same tool. It calls the wrong endpoint. It forgets a required field. It confidently does the one thing you told it never to do.

The instinct is to blame the model. That instinct is wrong.

The bottleneck moved

For two years the hard part was reasoning. Could the model plan, hold context, follow a chain of steps. That fight is largely over. The frontier models are good enough for real work.

So the failure moved. It is no longer that the model cannot think. It is that the model does not know how your specific tools work. Every session it rediscovers your wiring from scratch. Which API. Which field. Which of your two similarly named systems does what. It guesses. Often it guesses well. Sometimes it guesses wrong and acts anyway.

Plugging an agent into more tools does not fix this. It makes it worse. More surface area, more wiring, more to forget.

The forgetting tax

Name it, because naming it is half the fix. Every session an undocumented agent pays a forgetting tax. It re-derives how your systems fit together before it can do the actual job. That is wasted time on a good day.

On a bad day it re-derives it wrong.

Here is one from our own terminal. We were rotating a secret. Two keys had similar names. Nothing on the machine documented which key gated which system. The agent picked the wrong one, rotated it, and took down something that should never have been touched. A one page note would have made that a twenty second read instead of an outage.

That is the tax. It is quiet, it compounds, and occasionally it bites.

What a skill actually is

The fix is not a smarter model. It is a skill.

A skill is one Markdown file per tool. Short. It holds four things:

• What the tool is, in a sentence.
• The safety rules and the traps. What to never do, and where the sharp edges are.
• The interface. The shape of the thing, enough to start.
• The workflows. The two or three jobs you actually run.

Then it does the most important thing. It points the agent at the live source of truth. The API itself. The database. The real schema. The skill never goes stale because it does not try to be the whole manual. It is the map, and it tells the agent where the territory lives.

The agent reads the skill before it touches the tool. Every session. No rediscovery. No guessing.

Two kinds, and the second is the one people miss

There are two kinds of skill.

Vendor skills cover external tools. A payments API, a mail provider, a social platform. Useful, but the vendor already wrote the docs. You are just curating them for the agent.

Internal-wiring skills cover your own systems. The scripts nobody documented. The database only you understand. The way your own tools actually talk to each other.

The second kind is worth more. There is no vendor manual to fall back on. If it is not written down, the agent re-derives it every time, and your undocumented internal wiring is exactly where a wrong guess does the most damage.

Most people write the vendor skills and skip the internal ones. That is backwards.

Skills are guardrails, not just memory

Here is the part that matters most, and it is not about memory at all.

A raw API doc tells your agent what it can do. A skill tells it what it should do, safely, and where to look when it needs more.

The weight of the guardrails should match the blast radius of the tool. A read only reporting tool needs light rules. A tool that messages real people one to one, on your accounts, in your name, needs the guardrails to be the spine of the whole skill. What it can send. What always needs a human. What it must never do unprompted.

A model with a powerful tool and no skill is not clever. It is unsupervised.

How to start

Do not boil the ocean. Write one skill, for the tool your agents use most and break most often. One page. What it is, the traps, the interface, the workflows, a pointer to the live truth. Load it at the start of every session that touches that tool.

Then write the next one. The internal ones first, because those are the ones nobody else will document for you.

Bottom line

Your agent is not dumb. It is unsupervised and under briefed. The model was never the bottleneck. The knowledge was.

Give each tool a skill, and the flaky agent becomes a reliable one.

This is how we run our own agentic operation at Neon Gorilla. AI sales, search, and social, on your own accounts, briefed and guardrailed, not left to guess.