skill: draft auditing-alignment for spec/plan/code reconciliation

[intendednull]

Summary

New skill at .claude/skills/auditing-alignment/SKILL.md. Scoped to reconciling spec ↔ plan ↔ code at two checkpoints rather than the broad code-quality sweep that general-audit does today.

• Pre-implementation mode — spec written, no code yet. Catches shape conflicts (wrong async model, wrong state-management pattern, missing abstractions) before the implementer wastes time.
• Post-implementation mode — feature reported done. Walks spec → code, then code → spec, then reads the plan last to avoid biasing the comparison.
• Decision tree classifies every finding as code-update, spec-update, refactor, or discussion. The implementer's decisions can override the spec — surfacing those is the point.
• "All clear" is an explicit valid outcome. Red-flag table + hard rules push back on the urge to manufacture findings.
• Lessons-learned loop at the end edits this skill file in place — but only after human approval, never unsolicited.

Open questions

• Keep general-audit alongside this, retire it, or merge them? Scopes differ (broad sweep + GitHub issue filing vs. focused alignment audit, no issue filing).
• Is discussion too soft as a finding type? Should "implementer scope creep" be its own type?
• Should the trigger also fire mid-implementation if drift is suspected, or stay locked to the two checkpoints?

Test plan

• Run skill on a real recent spec/code pair, post-implementation, and check that the decision-tree classification feels right
• Run skill on a spec with a known shape conflict, pre-implementation, and check that it catches it
• Run a baseline subagent test (RED-GREEN-REFACTOR per writing-skills) once the shape is settled — deliberately skipped here since this is a first draft for iteration
• Decide what to do with general-audit

https://claude.ai/code/session_01CrqkjcE8US71mJQnvNwF4j

---

Generated by Claude Code