Skills as Organizational Knowledge Assets at the Automation Threshold: When Does Codified Know-How Become Reliably Executable by Agents?
Authors
Chircop, Claudio ; Röker, Niklas Finn Mats Philipp
Term
4. semester
Publication year
2026
Submitted on
2026-06-01
Pages
92
Abstract
Organizations are increasingly codifying work for AI agents into reusable “skills” built around a SKILL.md file, aiming to make know-how repeatable and delegable. This thesis treats skills as a codification and inscription layer between human expertise and machine execution and asks what is actually being inscribed in one public ecosystem. It reports a repository analysis of 120 publicly inspectable agent skills from the skills.sh registry, coding each artefact along five dimensions: knowledge type, codifiability, evaluability, task horizon, and governance. The corpus appears well codified but largely unverifiable: more than three quarters are well specified, yet only one in twenty-five includes means to verify outputs, and 87.5% ship no working tests. What gets codified depends chiefly on the kind of knowledge: procedural work codifies fully, while analytical and judgment-heavy work resists; this pattern does not seem driven by task length or complexity. The thesis argues these scores capture inscriptive effort rather than deployed reliability: the ecosystem yields well-specified procedural ambition whose results remain mostly unverified. It therefore frames the findings as claims about what publishers inscribe, not how skills perform, and considers what the artefact can and cannot tell an organization weighing adoption.
Organisationer forsøger i stigende grad at omsætte arbejde til agent-læselige “skills” bygget omkring SKILL.md, så knowhow kan genbruges og delegeres til AI-agenter. Denne afhandling behandler skills som et kodificerings- og inskriptionslag mellem menneskelig ekspertise og maskinel eksekvering og spørger, hvad der faktisk inskriberes i et offentligt økosystem. Den rapporterer en repository-analyse af 120 offentligt inspicerbare agent-skills fra skills.sh-registret, hvor hvert artefakt blev kodet langs fem dimensioner: videnstype, kodificerbarhed, evaluerbarhed, opgavehorisont og governance. Fundene viser et korpus, der er velkodificeret men i vid udstrækning uverificerbart: mere end tre fjerdedele er godt kodificeret, men kun én ud af 25 indeholder midler til at verificere outputs, og 87,5% leverer ingen fungerende tests. Hvad der kan kodificeres, afhænger primært af vidensarten: procedurepræget arbejde lader sig kodificere fuldt ud, mens analytisk og dømmekrævende arbejde gør modstand; dette mønster ser ikke ud til at afhænge af opgavens længde eller kompleksitet. Afhandlingen argumenterer for, at målene fanger inskriptiv indsats snarere end indsat pålidelighed: økosystemet producerer velbeskrevne, procedurelle ambitioner, hvis resultater i stor udstrækning forbliver uverificerede. Resultaterne skal derfor læses som udsagn om, hvad udgivere inskriberer, ikke om, hvordan skills faktisk performer, og diskuteres med henblik på, hvad artefaktet kan og ikke kan fortælle en organisation, der overvejer at adoptere det.
[This apstract has been generated with the help of AI directly from the project full text]
Keywords