AI discoveries outpace productization. New capabilities reach users before they’re integrated into reliable products, leaving users to build the workflows around them.

The danger is canonizing whatever happens to work. Useful intuition becomes prescriptive doctrine, and accumulated experience becomes an impediment to adaptation.

Musashi wrote, “You should not have a favourite weapon.”

Skills illustrate the problem. Their present utility establishes neither permanent superiority nor immunity from obsolescence. “Fat skills” don’t escape that scrutiny. Novel packaging doesn’t establish conceptual novelty, and procedural complexity requires evidence of corresponding benefit.

I predict the eventual obsolescence of all current scaffolding that compensates for missing AI capabilities. AI develops as trained ecosystems: models, tools, memory, and feedback. Useful behavior can be absorbed into the system, leaving external recipes redundant or actively restrictive. Prescribing yesterday’s workaround can prevent tomorrow’s system from exploiting its capabilities.

Economic and reputational attachment compounds the problem. When skills become someone’s business or identity, the incentives favor elaborating the method over acknowledging its obsolescence.

The only antidote to stagnation is flexible thinking: revisable priors, continually tested against changing capabilities. Intuition should accelerate experimentation without dictating its conclusions. Each model release warrants reassessing the assumptions embedded in our workflows.

Without that discipline, experience ossifies into doctrine. We become proficient at preserving techniques whose original justification has disappeared.

Test subtraction as rigorously as addition. Every technique must keep earning its place through demonstrated utility.

The model eats everything.