LLM Knowledge Base

Vibe coding lowers the floor of who can build software, and anyone can ship a working app in an afternoon. The ceiling stays a job: whether the result is correct, secure, and worth building, judged by someone who knows what correct looks like.

AI advice rots fast, but not evenly. Splitting notes into invariants and dated tactics keeps them trustworthy: claims about what stays human in one section, today’s model-specific tricks stamped with a date in the other.

Sixty years of census data record exactly one job eliminated outright by automation. The rest were transformed or multiplied, and 60% of today’s jobs did not exist 50 years ago. Whether a role gets augmented or replaced turns on whether it can absorb responsibility.

The mindset corpus compressed to doctrine: difficulty read as fuel, mistakes as the mechanism of improvement, motivation as startup fuel with defusion carrying the daily load, and gains judged by whether they compound.

For controlled-vocabulary authoring behind a fail-closed gate, Sonnet 5 becomes the default author at 0.6 times Opus 4.8's per-token price with near-identical first-pass gate rates, and Opus stays the orchestrator and the fallback for calls the gate cannot verify.

Every capability scores on two independent axes, Human and AI, each broken into five facets, and lands in one of four zones: Own, Augment, Delegate, or Low-leverage.

Two design books split into an Agent Track of 66 executable UI rules and a Human Track of 35 taste-and-judgment calls.

The dictionary's Story card becomes a single-character memory scene, carrying only the residue the Form card leaves: the arbitrary binding of components and the part of the sound the phonetic series does not predict.

The rules a memory scene must obey reduce to two axes that generate every card: thirty character situations crossed with forty-two memory levers.

Those two axes generate thirty-four card archetypes, the working set a future authoring agent selects from and extends.