differential-audit
Compare two implementations of the same thing — a port (R↔Python↔Stata), a reimplementation, a replication package, a refactor, or a new version against the old — so that agreement means something. Freeze inputs first, inventory every expected output, test the comparator itself, compare every channel (not just the headline number), and give each divergence a stable ID and a smallest witness. Use for cross-language parity, replication, upgrade/regression gates, or whenever "the numbers match" is about to license a claim.
pedrohcgs · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add pedrohcgs/claude-code-my-workflow --skill differential-auditAbout this skill
name: differential-audit description: Compare two implementations of the same thing — a port (R↔Python↔Stata), a reimplementation, a replication package, a refactor, or a new version against the old — so that agreement means something. Freeze inputs first, inventory every expected output, test the comparator itself, compare every channel (not just the headline number), and give each divergence a stable ID and a smallest witness. Use for cross-language parity, replication, upgrade/regression gates, or whenever "the numbers match" is about to license a claim.
differential-audit
This skill's instructions live in .claude/skills/differential-audit/SKILL.md.
Read that file now and follow it, with the files it points to. Read AGENTS.md
first if you have not already.
It was written for Claude Code. Running it from Codex, Antigravity or another agent:
- Arguments. Where it says
$ARGUMENTS, use what the user asked for. - Claude Code agents. Where a step hands work to an agent in
.claude/agents/(through theAgentorTasktool), do that step where it appears in the procedure: never later, and never skip it. A step that gates what follows still gates it (/commitruns its verifier before it commits). Read the agent's file for what the step checks. - Fresh context. If the step needs a context that has not seen your work, for example a claim
verifier that must never see the draft, run it in a separate session or worker that has not
seen the draft, if your tool can start one, and give it only what the step says to pass. If it
cannot, do the step yourself, mark every result it produced not independently verified, and
tell the user to repeat that check in Claude Code (
/verify-claimsfor citations) before relying on it. - Claude Code settings.
allowed-tools,model,effort,argument-hintanddisable-model-invocationin the source's frontmatter configure Claude Code. Ignore them. - Guards. Claude Code's guard hooks do not run for you. Follow the rules in
AGENTS.mdyourself; the git pre-commit gate and CI still apply to your commits.