Detailed Analysis
A systematic cataloguing effort spanning 2,392 Claude Code skill files has produced findings that sharply contradict the prevailing narrative around how Claude is being used in software development. The researcher, who spent three months assembling and categorizing the dataset, found that SAP-related skills account for 107 of the 845 curated entries — roughly 12.7% of the verified subset and four times larger than the next closest category, which is general database work at 26 skills. Frontend development, Python, and AI/ML — the topics that dominate discourse on platforms like X (formerly Twitter) and Reddit — each clock in at just 15 skills in the curated set. When Salesforce, ServiceNow, and Dynamics 365 skills are grouped alongside SAP, enterprise platform tooling collectively dwarfs every other vertical represented in the data.
The explanation for this disparity lies in the structural incentives facing enterprise platform consultants rather than any coordinated effort. ABAP debugging, Fiori UI migration, Salesforce Apex testing, and similar workflows are highly repetitive, deeply specialized, and largely underrepresented in general model training data — a combination that makes Claude Code skill files unusually valuable in those contexts. Because large enterprises frequently have compliance constraints that complicate the deployment of MCP servers or external integrations, lightweight markdown-based skill files sitting locally in a developer's environment represent a lower-friction alternative. The researcher specifically identifies this as a key reason enterprise practitioners adopt the skill file format at higher rates than their consumer-facing counterparts. Supporting evidence from GitHub repositories such as `secondsky/sap-skills` — which contains 35 production-ready skills spanning SAP BTP, CAP, ABAP, HANA, and Fiori — confirms that this community has developed structured, mature tooling rather than ad hoc experimentation.
The quality analysis embedded in the dataset is equally instructive. Of the full 2,392 catalogued files, only 789 pass a basic verification threshold covering syntactic validity, non-duplication, actionable content, and absence of prompt injection — a signal rate of roughly 33%. Three recurring anti-patterns account for much of the low-quality tail: skills that run over 3,000 words without actionable structure, generic persona prompts dressed up as skills, and files that are simply collections of viral prompting techniques repackaged as domain knowledge. The researcher's empirical finding that effective skills fall in the 200–800 word range maps to a functional constraint: beyond 800 words, a skill file begins competing with the actual task prompt for Claude's effective attention, degrading rather than augmenting output.
The broader implications of this data extend well beyond the immediate Claude Code ecosystem. It surfaces a recurring pattern in AI adoption curves, where the loudest online discourse — skewing toward developer-founders, startups, and modern web stacks — systematically underrepresents the enterprise segment that often accounts for the majority of real-world usage volume and economic value. SAP alone underpins transaction processing for roughly 80% of Fortune 500 companies, meaning the developer surface area for AI-assisted SAP work is enormous relative to its cultural footprint in AI circles. This dynamic has historically played out in enterprise software broadly, where products achieve scale inside large organizations before receiving proportional attention from the technology press or developer community.
For companies and developers building tooling, extensions, or marketplaces around Claude Code, the dataset functions as a corrective signal. The competitive analysis the researcher includes — covering Claude Code against Cursor and GitHub Copilot — likely reflects similar asymmetries, since both of those tools also receive coverage disproportionately weighted toward web development contexts. If enterprise platform verticals constitute the actual plurality of skill file demand, then product decisions optimized entirely around the vocal minority of solo SaaS developers may be systematically misdirecting investment. The planned v2 expansion of the dataset in July, with a particular call for enterprise-context skills, suggests the researcher intends to track whether these patterns deepen as Claude Code adoption continues to grow across organizational contexts.
Read original article →