Skills That Don't Guess

SKILL.md has become the most popular way to teach AI agents new abilities. It is simple and convenient. But it has one property that makes a big difference in business: whether, when and how a skill gets used is decided from scratch by a language model, every single time.

Michał WarchołSophona

An instruction the model reads its own way

The idea behind SKILL.md is very human: you write the agent an instruction the way you would write a note for a new colleague. The agent knows the titles of all its instructions, and when it decides one fits the task, it reads it and acts. For creative work — analysis, summaries, writing — it is a genuinely good solution.

But notice the word “decides”. Every time, the model settles on its own:

whether to use the instruction at allThe same question, worded differently, can skip it entirely.

how to interpret it“Check first, then send” is a request, not a guaranteed order of steps.

whether the task succeededIt is judged by the same model that did the work.

why it acted the way it didOften a question nobody can answer after the fact.

When an agent writes a report, you can live with that. When it issues documents, works inside a client's systems or runs a conversation by company procedure, you can't. There, what matters is not how the agent usually behaves, but what it does in the worst case.

Two models instead of one

At Sophona we don't ask one model to do everything. We combine two kinds of AI, each doing what it does best.

A small BERT-type model recognises the situation. It doesn't write and doesn't chat — it understands. It reads every message and, in a fraction of a second, works out what it is about: an offer, a document, a complaint, or just a conversation. Based on that, it selects the business rules the company has set for that situation — what the agent may do, what it must not do, what it should ask and what it should check.

A large language model, such as GPT, leads the conversation. It receives the rules that apply to this exact moment, so it doesn't have to guess which ones matter. It can focus on what it is great at: understanding people, reasoning and answering naturally.

~0.003 s That is how long it takes to recognise the situation. It happens locally, without an extra call to the large model — fast, cheap and the same every time.

The result is simple: it is the company, not the model, that defines which rules apply in a given situation. And because recognition is predictable, it can be tested before the agent ever meets a customer — type a sentence the way a customer would, and see which rules kick in.

A skill is a robot, not a request

The second difference is in how work actually gets done. At Sophona, an agent's skill is not a description to be interpreted. It is a robot — a designed workflow that performs the task the same way every time.

The language model decides that an invoice needs to be issued, and with what data. It does not decide how it is issued. That means:

order is enforced, not suggested — if a process requires signing in, choosing an account and confirming, the agent cannot skip a step;
the important questions come before any action — a robot can require the agent to settle something with the person first, and can ask for data or approval itself while it works;
“done” means done — a task counts as complete only when there is proof the action actually succeeded, not when the model says so.

What this gives a company

PredictabilityThe same situation triggers the same rules and the same way of working.
ControlRules and processes are set by the company, not by the model's hidden “intuition”.
VerifiabilityThe agent's behaviour can be tested before rollout and traced afterwards.
Natural conversationThe large model still does what it does best — talks like a person.

Sophona builds AI operators for businesses that talk like people and work like a well-designed process.

Deterministic AI Skills

こちらにも