audit-plan

Audit Plan Execution

Overview

Verify that every task in the session plan file was implemented correctly by comparing planned changes against actual state.

Core principle: Trust evidence, not assumptions. Check every task.

Announce at start: "I'm using the audit-plan skill to verify plan execution."

The Process

Step 1: Load The Plan File

1. Read the session plan file (path is in conversation context from the planning session)
2. Extract the task list and all file references

Step 2: Verify Each Task

For each task in the plan you MUST spawn subagents and send them in ONE message to:

1. Check file existence: Glob/Read files the task was supposed to create or modify
2. Check file content: Verify the changes match what the plan specified (exact code, structure, config values)
3. Check deletions: Confirm files or code blocks the plan said to remove are actually gone
4. Run verification commands: If the plan included commands with expected output, run them
5. Argue with the results: Understand the task goal and compare it with the actual results. If there are any discrepancies, you MUST report them.

Step 3: Report

Present a structured summary:

## Audit Summary

| # | Task | Status | Evidence |
|---|------|--------|----------|
| 1 | ... | Done / Partial / Missing | What was checked |
| 2 | ... | Done / Partial / Missing | What was checked |

### Issues Found
- [List any discrepancies, if none say "None"]

Status definitions:

• Done — All planned changes for this task are present and correct
• Partial — Some changes present, others missing or incorrect
• Missing — Task was not implemented

Step 4: Fix Missing Points If Applicable

If there are any missing points in the plan, you MUST spawn subagents and send them in ONE message to:

1. Fix the missing points: Understand the task goal and fix the missing points
2. Verify the fix: Verify the fix by following the guidance in the Step 2
3. Audit again: Go back to Step 2 and audit the plan execution again

Remember

• Read the plan file first, every time
• Check actual files, don't rely on memory of what was done
• Report evidence for each verdict
• If a task has no verifiable artifacts (e.g., "research X"), mark as Done if conversation shows it happened
• Don't fix issues found — just report them. The user decides next steps.