agent-primitives

Personal-style writing with explicit rules, advisory corpus tendencies, scoped persistent preferences and exact-final-text checks. Authenticated Claude/Codex runtime; optional whole human examples.

2 skills · 4 agents

Two read-only critics and one revising transformer for prose. prose-voice-critic: does this sound like the author, judged against their own corpus, with every finding required to cite how they write instead. prose-fidelity-critic: did a revision keep the facts, quotes, qualifications and claim strength of the original, judged against a deterministic scan that is authoritative on presence. prose-reviser: applies a plan of edits and emits a change log the fidelity critic then judges — out-of-plan edits impossible by construction. Never judges quality, never claims anything about who wrote something.

3 agents

Two read-only reviewers that run before an agent may declare a task done: verification-critic (reward-hacking and spec drift) and architecture-reviewer (structural decay). Optional Stop hook refuses to end a turn on red Tier-1 gates.

2 agents · hooks

Deterministic prose measurement: a catalogued scan for AI writing tells plus the cadence metrics a model cannot estimate about its own output. Register-aware, calibrated against your own corpus, and report-only — signals for an author about their own draft, never a verdict about who wrote something. Ships no rewriter, by design.

1 skill · 1 command