Wisdom is not stored as documentation. It is distilled into Seeds — compressed, generative heuristics that unfold into full reasoning frameworks on contact with a problem. Three families. One schema. Infinite application.
Wisdom is not stored as a SKILL.md. It is distilled into compressed, reactivatable patterns that unfold on contact with a problem — mirrors of how the human mind actually holds knowledge without structural collapse.
Authenticity is not a STYLE.css. It is the presence of Friction — the jagged edges of human cognition that LLMs smooth over into high-entropy slop. These seeds restore the resistance of a thinking mind.
Systems do not fail because they lack features. They fail because they lack Invariants. ArchSeeds are the physics of the design — hard constraints that ensure entropy remains manageable as complexity scales.
Most prompt engineering is behavioral — lists of rules, forbidden words, style guides. Seeds operate at a deeper layer. They give a system a coherent epistemic identity — a worldview from which correct behavior is derived, not prescribed. The model doesn't follow the rule. It becomes the kind of system that wouldn't violate it.
Every seed in the registry survives this decomposition. If it cannot be unpacked this way, it does not qualify.
| Context | CogniSeeds | LinguaSeeds | ArchSeeds |
|---|---|---|---|
| System Prompt | Epistemic discipline, reasoning depth, uncertainty handling | Voice authenticity, anti-sycophancy, friction injection | World logic, component contracts, failure assumptions |
| Code Review | Process/output distinction, invariant tracking | Documentation voice, comment clarity | SSoT, encapsulation, evolvability, observability |
| Content Editing | Epistemic honesty, source/claim mapping | Rhythm, concreteness, hedging removal, soul injection | Structure, argument load-bearing, logical removability |
| System Design | Mental model clarity, blast radius awareness | Documentation that survives the diagram going stale | All eight seeds apply directly |
| AI Agent Design | Reasoning identity, confidence calibration | Output voice, hallucination pattern removal | Interface contracts, failure-first design, observability |