Every time OpenAI ships a new model, there is a refrain: skills are useless, and more and more users are abandoning them. Is that right? Right and wrong.
The author grants the refrain's kernel: the skills being abandoned are mostly "generic" ones whose job was to patch model weaknesses — especially from the era before models were aggressively tuned for coding. As models got stronger and absorbed those abilities, the skills naturally stopped mattering.
But skills do not disappear; their effective scope just converges to the "medium-reusability" range: personal workflows (e.g. managing tasks in Notion and generating weekly reports), the use of a company's outdated in-house stacks, and a browser skill that carries your real cookie to visit anti-scraping sites. These are things the model "can't learn from its corpus alone"; the knowledge lives only in your head or your team's.
The author's software-engineering analogy: the skill layer maps to a SDK/toolkit, and the model maps to a programming language; the degree of reuse decides which layer a capability belongs in — arrays and maps belong in the language, invoice-parsing logic belongs in a toolkit. Translated to AI: a skill is "reusable enough to be worth wrapping, yet not so universal that AI can learn it from its corpus." And "when you download someone else's skill, it usually doesn't run well for you."
Conclusion: the stronger the model, the less skills can afford to be scattered. Universal skills get absorbed, and the retrieval that picks the right skill degrades as the pile grows; what deserves to be captured is only the small part you (or your team) alone know. That is why the author built Flint — treating skills as a personal asset: a single source of truth, tagging, deduplication, and on-demand deployment, rather than a giant "download-and-run" repository.
Sources · 参考来源