What a good AI skill really is
Not a personality trait or a decorative title, but a clear, tested capability: what it is for, what it needs, what it delivers and where it stops.
“Skill” sounds like a special talent. In work with AI it means something more sober: a repeatable capability you can describe, hand over and check. A good AI skill is therefore not a long role instruction or a decorative title. It describes where a capability is used, which information it needs, what it should deliver and when it must stop or ask.
The difference is practical, not academic. A vague role can be admired but not improved. A clearly bounded capability can be tested, corrected and sharpened a little the next time. That is how you know a chat history has turned into a reliable tool.
Capability, not personality
“Be an expert” leaves almost everything open. It says nothing about which task is solved, which material must be present and how a good result is recognised. A skill becomes useful only when it has a clear work area: checking inputs, reading a draft against fixed criteria or creating a structured hand-off.
The narrower a capability is described, the easier it is to test and improve. A personality can only be believed. A capability can be observed. That is why “checks an invoice for three mandatory fields” is a better starting point than “is a careful bookkeeper”.
Six fields turn a role into a tool
A skill does not need much to become useful: a task, the necessary inputs, a checkable output and clear boundaries. Six short fields are enough to make these parts visible and turn the capability from a mood into a tool.
Each field answers a single question. Together they form a profile a second person can read without you sitting beside them. That is the real test: can someone else apply the capability — or does it live only in your head?
The boundaries are the most important part
Most skills fail not because they can do too little, but because they try too much. That is why the last two fields — “Asks when” and “Must not” — are often more valuable than the first four. They set what may not be invented, when a question must be asked and when a person should decide.
This creates no artificial all-rounder, but a tool with visible edges. A skill that answers “I cannot say that with certainty” at the right moment is worth more than one that confidently papers over every gap. The boundary is not a weakness of the skill — it is its most important safety feature.
Skills are learned through cases
Rules alone remain claims until you test them against cases. So describe a normal case, a difficult case and a case where the skill must refuse or ask questions. Only these counterexamples show whether the rules hold or merely sound good.
A skill is good when you can understand and correct its result. The three cases are not bureaucratic overhead but the real definition: they describe the capability more precisely than any string of adjectives could.
| Case | What the skill should do |
|---|---|
| Normal case | clean input, clear task → delivers the output in the agreed format |
| Difficult case | incomplete or ambiguous material → marks the gap instead of filling it |
| Impermissible case | request outside the area → refuses or asks |
An example: the invoice-check skill
Take a small skill that checks incoming invoices for three mandatory fields: supplier, amount and due date. The “Can” field is narrow: exactly this check, no more. “Needs” is the invoice as text or file. “Delivers” is a short list with a result per field. “Checks” is the rule that each field gets either a value or a clear “missing”.
The two boundary fields make the skill safe. “Asks when” triggers if an amount is ambiguous or two due dates are named. “Must not” forbids inventing a missing date or rounding an amount. The same skill, without these two fields, would quietly fill gaps — and that is exactly what you do not want with invoices.
The skill profile
Keep the profile short. Six fields, one or two sentences each. Whatever does not fit here usually belongs in a second, separate skill.
# SKILL
**Can**
Exactly this one recurring task …
**Needs**
These inputs must be present …
**Delivers**
This output in this format …
**Checks**
A good result is recognised by …
**Asks when**
… is missing or unclear
**Must not**
Invent, round or silently complete …A good skill does not make an AI human. It makes a recurring task clearer, safer and easier to check — and that is the point. When you can describe your capability in six fields, you do not have a magical system, but a reliable tool.
Worksheet: Define a small skill
Choose a recurring task from your field and describe it with the six fields. Keep it short: one or two sentences per field.
Choose one capability. State one repeatable task — not “help me with everything”, but a clearly bounded activity.
Define inputs and output. What must be present, and what should result at the end, in which format?
Set boundaries. When must the skill ask or stop? What may it never invent?
Test three cases. Write a normal, a difficult and an impermissible case — and note what the skill should do in each.
Both working materials for this article — the topic overview and the worksheet with a reflection space to fill in — are available for download here:
● Members only
Read the full article and download all files with a membership.
Unlock full article + downloads → Subscribe0 comments
No comments yet — be the first.