← ClaudeAtlas

verifierlisted

Verifies this session's own finished work before commit and fixes what it finds. Checks completeness first, to catch dropped tasks and unrequested additions, then runs parallel reviews for shape (is each touched unit the least code that does the job), logic and tests, and security, has a fresh agent try to refute the defects found, and applies the fixes.
viettdann/arclight-kit · ★ 0 · AI & Automation · score 65
Install: claude install-skill viettdann/arclight-kit
# Verifier: Completeness Check and Code Review Arguments: $ARGUMENTS Phase 0 confirms every planned or requested task was done; Phases 1-2 review the changed code and fix it. Completeness runs first even when the user asks only for a code review: tasks get dropped silently through context truncation, interruptions, or oversight, and reviewing the code that exists never surfaces the code that does not. ## Phase 0: Completeness Verification Establish what was supposed to happen, then check it against what actually changed. Run this phase yourself: the conversation is already in your context, and a plan file is short. ### Step 1: Identify the source of truth 1. **Plan file**: most reliable, not subject to truncation. Use the path in the arguments, else a plan file named in this conversation. Without either, search for `plan.md`, `PLAN.md`, `TODO.md`, `tasks.md`, `checklist.md`, `*.todo`, and design docs under `docs/plans/`, and use one only when it describes this session's work (it names files this session edited or that `git status --short` lists); with more than one match, ask which. Never take tasks from other files that happen to hold checkboxes: handoff notes, READMEs, changelogs. 2. **Conversation history**: ad-hoc requests, mid-stream changes, and corrections that never reached a plan file. With both (the common case), the plan file is the baseline and the conversation carries additions and modifications; a conversation change to a plan task overrides the plan's wor