prompt-quality-auditor
Prompt quality auditor: audit every prompt for clarity, safety, and injection resistance. Read-only. Trigger: /prompt-quality
SKILL.mdskills/ai-ml/prompt-quality-auditor/SKILL.md
GitHub/prompt-quality
Prompts are the most critical code you never review. This skill finds quality issues before they cause bad outputs.
What this is for
- Prompts without output format specification (model guesses format)
- Missing safety instructions (model complies with harmful requests)
- User input directly in prompt without sanitization (injection)
- Read-only skill. No prompt modifications.
PROTECTION RULE - never ~/.claude/
Read-only skill. Guard required if write mode added later.
What You Must Do When Invoked
During analysis, assign a confidence level to each finding: proven (confirmed by evidence), likely (strong signal, needs review), or suspected (weak signal).
If /prompt-quality -help or /prompt-quality -h (without further arguments)
is invoked: output the ## Usage section unchanged and stop.
Otherwise follow these steps in order, skipping none.
Step 1 - Help check
If invoked with -help or -h, output the ## Usage section unchanged and stop.
Step 2 - Clarify target
Clarify -ProjectDir. Get confirmation.
Step 3 - Scan
& "<SKILL_DIR>/scripts/prompt-scan.ps1" -ProjectDir "<path>"
Step 4 - Classification
Read each prompt:
- Structured: has output format spec + constraints
- Semi-structured: has some guidance
- Open-ended: no constraints on output
- Injection-susceptible: user input directly in prompt
Step 5 - Write report
File prompt-quality-report.md in current working directory:
- Summary - prompts by classification.
- Prompt table - injection-susceptible first. Per prompt: file, line, type, has format spec, has safety instructions, risk, recommendation.
- Open questions.
Step 6 - Summarize
State report path, highlight prompts that could produce harmful or unreliable outputs.
Usage
/prompt-quality # interactive
/prompt-quality <dir> # scan project
/prompt-quality -help