Commit

Mit Instruction Harness/Agent Skill for grai (open source) today: https://github.com/adrianpeticila/graiadrianpeticila/grai TL;DR: 134,522 LinkedIn posts syntactic audit 2023 to 2026 to understand why after removing all banned words, the LLM-generated drafts still feel synthetic. Reduced AI verbs by 88.1%. Replaced by structural symmetry. 1. Increased "here's why" 7.4x (from 1.3% to 9.6%) 2. Increased "it's not X, it's Y" contrastive negation 4.9x (from 1.2% to 5.9%) 3. Increased arrow bullet cascades 20.5x (from 0.6% to 12.3%) 4. Sentence burstiness (CV = sigma/mu) reduced 52.1% (from 0.48 human variance to 0.23 flat 15 words). I can't solve burstiness with prompts because RLHF models use style as soft hints. So I decoupled generation and verification. 5 files markdown harness (MIT). + 6 rules deterministic (regex) + CV>=0.32 linter in vanilla JS. One fun browser thing when I was developing offline workbench (https://grai.page). Browser doesn't allow me to use crypto.subtle.digest("SHA-256") from WebCrypto on file:// protocol. It's not secure context. So I wrote pure JS fips 180-4 SHA-256 to be able to hash offline local voice profiles without localhost server.