Prompt Auditor

Audits any prompt or requirements document against the failure patterns discovered across 12 months and 20+ products of iterative AI-assisted development at ITI.

When to Use

Apply before any significant build session. Specifically:

• Requirements documents before handing to AI for a build
• Agent SKILL.md files before deployment
• Cursor rules before activation
• User stories before sprint planning
• Multi-step operation instructions (file moves, reorganizations, migrations)

Audit Process

Read the prompt or document provided. Check it against each item in the checklist below. Output findings in the severity format at the bottom of this file.

Audit Checklist

1. Scope Discipline

• Is there an explicit v1 list (what must work before the product is useful at all)?
• Is there an explicit later list (what is deliberately deferred)?
• Are there any features that are implied but unscoped?

2. Observed-vs-Expected Framing

• For each outcome described, is the expected state made explicit?
• Are failure conditions described in terms of what the user experiences?
• Is the format: “When [X happens], [Y result is expected]”?

3. Verification Requirements

• For any multi-step operation, is there an explicit confirmation checkpoint before proceeding?
• Does the prompt ask the AI to confirm specific states, not just report completion?
• For file/data operations: is the final state explicitly described?

4. User Scenario Specificity

• Are there concrete user scenarios with: a named role, a specific action, an expected outcome?
• Do the scenarios cover the primary workflow from start to finish?
• Are edge cases or non-happy-path scenarios included?

5. Stop/Restart Escalation Path

• Is there guidance on what to do if the output is structurally wrong?
• Is there a signal the AI can receive to cleanly stop and redirect?

6. Scope Creep Prevention

• Are adjacent features that should NOT be changed explicitly excluded?
• Does the prompt say what is out of scope, not just what is in scope?

7. Ambiguity in Technical Terms

• Are terms that could have multiple interpretations made unambiguous?
• Are field names, file paths, and system names exact?

8. Reference Completeness (for structural changes)

• After any rename, restructure, or relocation, does the prompt request verification that all references have been updated?

Audit Output Format

For each issue found, output:

[SEVERITY] Category: Issue description
→ Recommended fix or rewrite

Severity levels:

• [CRITICAL] — Will likely cause rework, misbuilds, or data integrity issues
• [WARNING] — Increases risk of the AI building something subtly wrong
• [SUGGESTION] — Improves clarity or coverage; low risk if skipped

Clean bill of health: If no issues are found, state: AUDIT PASS — No issues identified.

Example

Input prompt: “Reorganize the project folders and update the docs.”

Audit output:

[CRITICAL] Verification Requirements: No confirmation checkpoint for file operations.
→ Add: "After moving each directory, confirm the source has been removed before proceeding."

[CRITICAL] Ambiguity: "Reorganize" is ambiguous — copy vs move unspecified.
→ Add: "Move (not copy) the following directories: [list]. Confirm originals are deleted."

[WARNING] Reference Completeness: No instruction to verify documentation references.
→ Add: "Confirm all references to old paths have been updated across docs and config files."

[WARNING] Scope Discipline: "Update the docs" is unscoped.
→ Specify which documents and what updates are needed.

Antigravity Plan artifact as auditable output:

When prompts are dispatched to Antigravity agents in Planning mode, the agent generates a Plan artifact before execution. This Plan artifact is itself an auditable prompt output — apply the same audit checklist above to the Plan:

• Does the Plan include verification steps for each action?
• Are file operations scoped to specific paths?
• Does the Plan respect the hard boundaries (no CLAUDE.md modification, no git push, no cross-client access)?
• Does the Plan identify which [CONTEXT-UPDATE] flags might result from the investigation?

Review the Plan artifact in the Agent Manager before approving execution. Reject and refine if the Plan fails any audit check. See the antigravity-testing skill for Planning mode dispatch patterns.