Overmind
The private vault where AI infrastructure patterns earn their place — never invented, only sweep-promoted from work that survived contact with reality.
Overmind is my private R&D vault for AI infrastructure — the patterns, references, and experiments that survive contact with real work.
It's not a product. It's the place where techniques that actually worked get written down once, so I can reuse them across every project, every business, every client.
Ideas land in an inbox. A scheduled job sweeps every Friday and proposes which ones graduate to durable patterns or reference docs. Things that don't earn their place get cut.
The discipline rule is simple: if removing every business-specific name from a pattern leaves it useless, it doesn't belong in Overmind. Universal only.
A markdown-only repo at ~/Code/overmind/ with four top-level dirs: patterns/ (durable techniques), reference/ (lookup knowledge), experiments/ (in-flight tests), and inbox/ (sweep proposals awaiting review).
A launchd-triggered Friday 18:00 cron runs scripts/sweep.js, scans recent activity across project repos for promotion candidates, scores them, and writes proposals to inbox/ for human review via the /sweep skill.
A pre-commit hook blocks files containing business-specific names from being committed unless explicitly allowlisted. Mechanical defense, not willpower.
Three tiers feed each other. Tier 1 is the hand-written manifesto. Tier 2 is cross-project Claude behaviors loaded into every session via @import. Tier 3 is per-project dream-extracted rules. Promotion is a two-path OR: rules graduate on within-project elevation (Claude behavior, not business logic) or cross-project frequency (2+ projects exhibiting the same pattern).