jhoblitt

Skill Review Claude — Development skill for Claude Code

Development community

Claude Code plugin marketplace: skill-review — a duplication and drift gate for skill, agent, and plugin prose.

How to install Skill Review Claude

This entry records only its repository, not the path inside it, so there is no exact command to give. Open jhoblitt/skill-review-claude and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Skill Review Claude does

Claude Code plugin marketplace: skill-review — a duplication and drift gate for skill, agent, and plugin prose.

Alternatives in Development

Thesis Drift — 投资论文漂移检测：分清事实变化与措辞变化 15.9k ★
Icpg Drift — /icpg-drift — Show All Drift 705 ★
Stop Review Gate — You are reviewing uncommitted changes before the author finishes their turn 664 ★

README

skill-review-claude

A [Claude Code plugin marketplace](https://code.claude.com/docs/en/plugin-marketplaces) for skill and plugin authors. One plugin, `skill-review`, carrying a four-axis review gate for instruction prose: skills, agents, canon and reference files, marketplace repos.

Instruction files fail in four ways code does not, and each axis enforces one rule against one of them:

axis	rule	what it catches
drift	every rule has exactly one normative statement; every other mention is a pointer to it	a rule restated in a second place, two renderings that have drifted apart, a pointer whose target moved, a rule with no normative home
concurrency	a procedure's control flow matches the dependency structure of the work it describes	independent work walked one item at a time, subagents spread across turns so they run serially anyway, a collect-all barrier that earns nothing — and unsafe fan-out: parallel writers without isolation, order-dependent work, unbounded sets with no cap
token usage	every operation runs on the cheapest mechanism that can do it correctly	a supervisor tiered below the decisions it must make, mechanical work on a frontier model, code the skill could ship but makes an agent rewrite, a run of tool calls one call could carry, prose billed on every trigger that only some invocations need
security	content crossing a procedure's trust boundary carries no authority over what it does	untrusted content in a prompt unfenced or behind a forgeable fence, prose that obeys instructions inside it, a step reading hostile input with a shell or an outward channel in reach, untrusted content promoted to execution, secrets on wide channels

Every one of these reads fine on the page — a restatement fails only once a copy moves, prose carries no price tag at all, and an injection surface looks like plumbing until someone writes the content that uses it — which is why they survive the review