brokenrobot-xyz

by Tamas Mezei

A suite for authoring Claude Code artifacts: reviewing-claude-skills and reviewing-claude-subagents each review their artifact in two passes — structure first, stopping at a structural verdict when the shape fails (fit-for-purpose leads the subagent gate), otherwise a full sweep ending in a computed verdict that honors the artifact's recorded waivers; improving-claude-skills and improving-claude-subagents loop the matching review autonomously — review → apply every blocking finding → commit → re-review — until the verdict is acceptable, the findings plateau, or the round cap is hit, asking only one question: the intent brief at kickoff. The subagent loop ends instead of converting when a review says the subagent should be a skill.

by Tamas Mezei

Authors Conventional-Commits commits and enforces them with a deny-hook, resolving the commit vocabulary from the host project's .brokenrobot-xyz/commits.json with vanilla Conventional Commits defaults.

by Tamas Mezei

A suite of frontend skills, starting with updating-dependencies: detects and categorizes outdated npm packages, researches the bumps the user selects with a dedicated subagent, and applies only the bumps the user approves — keeping each entry's existing version prefix and attributing new audit advisories to the update. Never commits.

by Tamas Mezei

Supplies criteria groups B–G for grading a Markdown prompt that steers Claude — model-specific prompting, general prompting, hallucination guards, output consistency, injection defenses, and prompt-leak defenses. Returns the criteria; the caller scores against them.

by Tamas Mezei

Revises agent-facing prose — skills, agent definitions, specs, and technical docs — so an agent cannot read a sentence two ways, applying twelve conventions adapted from ASD-STE100 Simplified Technical English.