Verification skills for Claude Code.
Trust nothing, verify everything. Assay provides post-implementation audit and verification skills that catch gaps between what was specified and what was built.
| Skill | What it does |
|---|---|
/design-audit |
Verifies implementation matches a design spec — catches specified-but-not-implemented gaps |
/pre-derive |
Independently derives expected requirements before seeing a plan, preventing anchoring bias |
/completion-review |
Post-ship codebase impact audit — finds redundant code, adoption opportunities, documentation drift |
/validate-plan |
Adversarial plan validation with binary PASS/FAIL verdicts and red flag scanning |
/verify-implementation |
Mechanical plan-vs-code compliance — counts tests, matches deliverables, checks scope |
git clone https://github.com/marcus-hooper/assay.git ~/Source/assay
cd ~/Source/assay && ./setupThe setup script symlinks skills into ~/.claude/skills/ where Claude Code discovers them automatically.
cd ~/Source/assay && ./uninstallThese skills are built on a workflow methodology where:
- Plans are contracts — the plan defines exactly what must be implemented, no more, no less
- Verification is independent — derive expectations before seeing the implementation to avoid confirmation bias
- Binary verdicts — PASS or FAIL, not "mostly complete" or "minor issues"
- Count everything — if the plan says 8 tests, count 8
it()blocks in the spec file - Trust nothing — don't trust the plan's claim it covered the spec, don't trust the implementation report's claim everything was done
Early development. Skill definitions are scaffolded with workflow outlines. Full prompt implementations are in progress.
MIT