Skill files: the next AI skill after prompting

The next skill after prompting

Prompting was the AI skill of the past few years: the right question, the right context, the right refinement. The reflex says vibe coding comes next – software built in dialogue until the result fits. That works for prototypes. As a team skill it doesn't scale: what comes out of it depends on the person who ran the dialogue.

The more useful skill is less glamorous: writing skill files. A skill file is a one-page, task-specific operating manual – precise enough that an AI executes the task without you sitting next to it. The weekly report, ticket triage, release notes: described cleanly once instead of re-prompted every week.²

Behind this sits the same shift that carries the whole Teklens cycle: knowledge moves out of heads into shared, versioned systems. A prompt lives in one person's head and chat history – when they resign, both are gone. A skill file sits in the repository or on the shared drive, has a history and belongs to the team.

What a skill file is – and what it isn't

Technically a skill file is a markdown file, often plainly skill.md: name of the task, trigger condition, steps in a fixed order, a definition of done, reference examples. Anthropic has formalised the pattern as Agent Skills – folders of instructions and resources a model loads once the task fits.¹

Three boundaries keep the term sharp. A skill file is not a prompt: it isn't invented per session, it is maintained. It is not documentation for humans: it is written so a machine can work from it – though humans read it with profit. And it is not a model feature: it works across tools, because it is nothing but text.

The test is simple: can a colleague – or an AI – execute the task from the file alone, without asking you? If not, the file is missing something, not the AI.

Prompt, skill file, spec: three ranges

The three tools differ not in style but in the reach of what they capture.

The prompt. Covers one session and one person. It is fast, flexible – and volatile. Nobody can reconstruct why last week's result looked different.

The skill file. Covers one recurring task and one team. It is the smallest versioned artefact in the system: one page, one owner, one history. When the process changes, the file changes – not the memory of five people.

The spec. Covers one product change. It is the contract of the define phase: requirements, constraints, acceptance criteria – agreed before an agent builds. The skill file is its small, operational counterpart: the same discipline, applied to the daily routine instead of the product.

What skill file and spec share is the core of the define phase: intent is agreed in writing before the machine runs. If you can write specs, you can write skill files – the exercise is the same, only smaller.

Your first skill file in an afternoon

Pick a task with a stable shape: the same structure on every repetition, a clear good-bad judgement, at least weekly frequency. The weekly report to management, triage of incoming tickets, the first draft of release notes. No strategy work, no one-offs – there the judgement sits in the context, not in the process.

Then answer five questions in writing: what is the task? When is it used? Which steps, in which order? What does done look like? And which real outputs should the result be measured against? The fifth question matters most: attach three good examples and one bad one – with one sentence on why it is bad. Examples calibrate more sharply than any description.²

You don't have to write the draft yourself. Give an AI your five answers plus the examples and let it draft the skill file – it knows the format. Then the real test: run the task on real work, compare against your bar, fix the file rather than the output. Two or three rounds and the file holds.

From one file to a team library

One skill file is the start, not the goal. Ten recurring tasks of one function – product, support, marketing – become a library: one file per task, exactly one owner per file. The owner is not a formality. It is the answer to who changes the file when the business changes.

Distribution is built in by now: skills can be packaged as plugins and rolled out to the whole team – updated centrally instead of copy-pasted through chat channels. For a B2B team in the DACH region this is the real lever: new hires get the way of working on day one, audit questions (“How is this report produced?”) have an answer you can show, and quality no longer depends on who happens to be around.¹

Skill files decay – people don't

A skill file doesn't learn. The business changes – the pricing logic, the tone, the report's audience – and the examples in the file go stale quietly. Without an owner nobody notices until the output misses. The mechanism is the same as with models in production: decay is the default, not the exception.

The fix is unspectacular: you learn, then you teach the file. A fixed review rhythm per quarter, a fresh example after every relevant change, a look at the places where users regularly rework the output.

Which is exactly why a skill file doesn't replace your best people. It captures their judgement in executable form – but the judgement itself still originates with them. The file doesn't make good people redundant; it makes their standards the team's standard.

Frequently asked questions

Recommendations

• Start with one task, not an initiative. A skill file for the weekly report, built in an afternoon, beats an “AI enablement programme” in quarter three. Start small, test on real work, then expand.
• Use the five questions as the frame: Task, trigger, steps, definition of done, reference examples. If one answer is missing, the task isn't skill-ready yet – or it belongs in a spec.
• Examples before adjectives. Three good outputs and one bad counter-example say more than “professional and concise”. Update the examples when the standard changes.
• One owner per file, visibly named. Ownerless skill files are stale skill files. Set the review rhythm before the library grows.
• Version skill files like specs. A repository or shared drive with history – not a private folder. The file belongs to the team; otherwise it is just a longer prompt.

Scope & caveats

• Skill files suit tasks with a stable shape. Strategy work, discovery and one-offs with a high context share don't belong in them – no file replaces the judgement in the room there.
• Formats differ by tool (Agent Skills, system prompts, context files such as CLAUDE.md). The principle – a maintained, task-specific manual – transfers; the technical details in this piece follow Anthropic's documentation.
• Time frames such as “an afternoon” describe the order of magnitude for a well-chosen first task – not a promise for complex processes with many exceptions.

Sources

Every external figure and quote in this piece – linked so you can verify it.

1. 1.Anthropic Docs, «Agent Skills» ↗ – Skills as folders of instructions and resources; packaging and distribution as plugins.
2. 2.Entrepreneur (2026), Essay zu Skill Files als nächster KI-Kompetenz ↗ – Core ideas: skill.md over vibe coding, the five questions, the team library and the owner principle.

Keep reading in the PM Lab

Related deep dives – from the same pillar and the adjacent phases.

Matching use cases from the library

From the article straight into practice: these use cases put the concepts to work with Teklens.