Every team eventually has a folder of prompts that nobody uses consistently. The same paragraphs get pasted with small variations, the output drifts, and the person who wrote the original leaves. A skill is the alternative, and it is not always the right one.
This page is the decision rule we use, with the writing tasks where the answer differs.
The short answer
If you would paste the same instruction a second time, make it a skill. If the instruction changes every time — a different audience, a different format, a different tone — it is a prompt, and turning it into a skill adds ceremony without adding consistency.
What a skill actually is
A skill is a folder with a SKILL.md: when to use it, what it does, what it must not do, and how to verify the result. Agents load it on demand when the trigger matches. The Skill Creator skill exists to write them correctly.
The key line is the trigger. A skill that does not state when it applies gets loaded at the wrong moment, which is worse than not having it.
Three writing tasks, three answers
| Task | Prompt or skill | Why |
|---|---|---|
| Rewriting a paragraph in a client's voice | Prompt | The voice changes per client; the instruction does not repeat |
| Every internal announcement | Skill | Same structure, same reader, produced weekly |
| Drafting a doc from an outline | Skill | The process is fixed; only content varies |
| A one-off landing page | Prompt | No repetition, no standing rule to enforce |
The pattern: frequency and sameness, not importance. A critical task done once is still a prompt.
The costs of a skill
Skills are not free. Each one adds context when it loads, and a library of twenty skills competes for the same window as the actual work.
They also create a maintenance obligation. A skill that encodes outdated format rules will keep producing outdated output long after the rules changed, and nobody will notice because the output looks consistent.
That is the failure mode to guard against: consistency is the benefit and the trap.
The costs of a prompt
The mirror image. A prompt library decays silently: versions fork, nobody knows which is current, and the best version lives in one person's notes. The cost is hidden until that person is on holiday.
If you have more than about five prompts for the same job, that is a signal the job wants to be a skill.
How to migrate a prompt into a skill
Four steps, and the first is the one people skip:
- Collect the variants. Diff the prompts you actually use. The differences are the parameters; the common text is the rule.
- Write the trigger. One sentence saying when to load this skill and when not to.
- Write the constraint. What the output must not contain — the failure the prompt was compensating for.
- Write the check. How a person verifies the result without reading the whole thing.
The fourth step is what separates a skill from a saved prompt. A skill that cannot be checked will produce confident output nobody validates.
Where writing skills pay off most
Internal communication is the clearest case: the same audience, the same structure, repeated weekly. The Internal Comms skill exists because that job is repetitive and shared.
Document drafting is the second: the process of agreeing an outline, drafting sections, and revising against criteria is stable even when the content is not. That is precisely the shape of a skill.
A migration example
A team writes a weekly changelog. The prompt has drifted into six versions; the structure is the same every week: what shipped, why it matters, what is next, one screenshot reference.
That is a skill. The trigger is "weekly changelog for the product", the rule is the four-section order, the constraint is no marketing adjectives, and the check is that the first section names a shipped item with a date.
Compare with a launch announcement for a new pricing tier. It happens twice a year, the audience differs, the tone changes. That is a prompt, and a skill for it would be a rule that gets loaded once and then rots.
The decay test
Here is the question that catches most bad skills: if we changed the format next month, who would notice? A skill with no owner and no check keeps producing the old format confidently. That is why the check step matters more than the trigger.
A skill you cannot verify in under a minute is a liability, because the consistency it provides will outlive the rules it encodes.
Frequently asked questions
Are skills better than prompts? No. They are better for repeated tasks with stable rules. Prompts are better for one-off tasks where the instruction is bespoke.
How many skills should a project have? As few as possible. Each one costs context when it loads; a dozen unrelated skills is a sign of unclear scope.
Can a skill call another skill? In practice yes, through references. Keep the chains short — a skill that depends on three others is hard to reason about.
Last checked: 9 November 2026.
Related: writing tools after generic voice, publishing with an agent, and skills.