Domain-specific skill libraries are the real agent moat, not core infrastructure

Comparing Claude (14 tools), Microsoft Copilot (2 tools), and Shortcut AI (1 generic execute_code tool with a rich TypeScript API) in Excel reveals that core agent infrastructure is replicable: “An elite team can build Claude’s 14-tool architecture with auto-verification in a few months.” The real competitive advantage lies in domain-specific skill libraries — Shortcut’s marketplace of LBO modeling, ASC 805 compliance, and bankruptcy workflows embodies “domain expertise that took years to develop.”

Garry Tan describes how such a library actually accumulates, and names the unit. The move is to “skillify” any hard-won result: “you do some feat of strength and then you turn it into a markdown file plus code plus tests that can be reused and put into a cron job.” His framing of what you now own is economic — “a markdown file is an employee,” one that “will do the job perfectly every single time and will do it as many times as you want.” The accumulation is a ratchet rather than a stock, because failures convert into permanent capability: the first attempt “is going to be bad,” but correcting the agent in plain language turns the trace into a skill file, and “anytime it screws up in a future case, it’s just a bug fix, and then it’s there forever.” He cites companies going “from zero to 15 mil ARR in about 4 months with like two or three people and a few hundred skill files.” The compounding has a maintenance cost, though — see Once skill files accumulate, provenance becomes the binding constraint, where the library grows large enough that its own facts start to disagree.

This vindicates Markdown skill files may replace expensive fine-tuning at the market level: skills aren’t just a technical architecture choice but a moat strategy. For data agents specifically, Context layers supersede semantic layers for agent autonomy shows how the moat extends beyond skill libraries to encompass canonical entities, identity resolution, and governance — a full context layer that traditional semantic layers can’t match. It also reinforces Context is the product, not the model — “the model matters less than the tools,” and the tools matter less than the accumulated domain knowledge they encode. The defensibility ladder climbs from model (commodity) to tools (months to replicate) to domain skills (years to accumulate) to user data (irreplaceable). This matters even more because Frontier companies absorb every useful agentic pattern into their products — if core infrastructure gets absorbed into foundation model products, only accumulated domain expertise remains defensible. Beyond static skill libraries, Proprietary feedback loops create moats that widen with every interaction shows that dynamic learning from usage creates an additional moat layer that competitors can’t replicate — skills encode what you knew, feedback loops encode what you’re learning.

Tan pushes the same argument down to the individual and makes it a question of ownership rather than defensibility: Own your skill files, because if you don't, your job becomes a skill file. If a skill file is cognition extracted and made executable, then the accumulated library is a moat regardless — the only open variable is whose repo holds it, and a worker whose judgment compounds in the employer’s repo has, in his phrase, “an extraction” rather than a career.