Claude Skills by benyoskovitz

Grade the current diff against the task rubric written at .rubric.md before implementation. Grading runs in a fresh, independent sub-agent so the agent that wrote the code is not the one certifying it. On PASS, clears the change for /pre-commit and archives the rubric. On FAIL, outputs a re-plan directive and stops, never patches in place. Run after Execute, before /pre-commit.

Run a two-lens audit on the current change, security/bugs in one pass, and conformance to the project's own quality rules in another. Use on demand after a non-trivial change that touches sensitive surfaces (auth, database queries, user input, file uploads, encryption) or when you want a deeper read than /pre-commit. Reads the diff and the project's CLAUDE.md directly; no server, no setup.

Removes AI-writing tells from any draft: em dashes, buzzwords and corporate filler, throat-clearing transitions, manufactured candor, forced rule-of-three, "it's not X, it's Y" constructions, reveal-bait setups, and other statistical-average patterns. Use whenever writing or polishing prose, and always as the final pass on anything written in a specific voice. Also invoke directly on any draft ("run the humanizer on this," "de-AI this"), including text a human wrote or pasted from another too...

Pre-commit quality gate. Scans the staged diff for known anti-patterns and security issues, runs the test + type-check suite, and blocks on any BLOCK-level hit. Run before every commit. Reinforced by the PreToolUse hook in hooks/settings.json.

Use when a bug, test failure, or unexpected behavior resists the first obvious fix, before proposing another one. Enforces root-cause investigation, one hypothesis at a time, and a hard stop after 3 failed fix attempts.