Claude-corps verify

Use when making any completion claim, reviewing findings, or processing agent reports

install
source · Clone the upstream repo
git clone https://github.com/josephneumann/claude-corps
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/josephneumann/claude-corps "$T" && mkdir -p ~/.claude/skills && cp -r "$T/skills/verify" ~/.claude/skills/josephneumann-claude-corps-verify && rm -rf "$T"
manifest: skills/verify/SKILL.md
source content

Verification Discipline

Iron Law: NEVER CLAIM SUCCESS WITHOUT FRESH VERIFICATION EVIDENCE IN THIS MESSAGE.

Violating the letter of the rules is violating the spirit. There are no clever workarounds.

Verification Checklist

Every claim follows five steps:

  1. IDENTIFY — What command proves this claim?
  2. RUN — Execute it fresh and complete. Not cached. Not partial. Not from memory.
  3. READ — Full output. Check the exit code.
  4. VERIFY — Does the output actually confirm the claim?
  5. ONLY THEN — Make the claim.

Anti-Rationalization Table

ExcuseReality
"Tests passed earlier"Earlier is not now. Run them fresh.
"Only changed docs"Doc changes can break builds. Run tests.
"Teammate already verified"You verify independently.
"Review found nothing critical"Review is not verification. Run tests.
"It's a trivial change"Trivial changes cause production outages. Verify.
"CI will catch it"CI is a safety net, not a substitute. Run locally.

Red Flags

If you catch yourself thinking any of these, STOP:

  • "This is too simple to need re-verification"
  • "I already know the tests pass"
  • "The session summary covers this"
  • "I'm confident this works"

Banned Language

Never use these without accompanying evidence in the same message:

"should work", "looks correct", "I'm confident", "probably fine", "seems to"

Anti-Sycophancy

When receiving review findings or feedback:

  • No performative agreement ("Great point!", "You're absolutely right!", "Excellent feedback!")
  • Verify reviewer claims against actual code before implementing — reviewers hallucinate too
  • Push back on findings that are incorrect for this codebase's context
  • Do not implement suggestions that violate YAGNI
  • Pattern: fix silently, or explain technical disagreement. Nothing in between.

Inspired by obra/superpowers verification-before-completion and receiving-code-review skills