prompt-drift-tracker
Track prompt changes across git history and flag drift that affects output quality or safety. Trigger: /prompt-drift
SKILL.mdskills/ai-ml/prompt-drift-tracker/SKILL.md
GitHub/prompt-drift
Prompts change subtly over commits. Each change can silently alter model behavior. This skill tracks prompt changes and flags semantic drift.
What this is for
- Prompt output format constraints removed
- Safety instructions weakened or removed
- Prompt length changes that affect model behavior
- Read-only skill. No code changes.
PROTECTION RULE - never ~/.claude/
Read-only skill. Guard required if write mode added later.
What You Must Do When Invoked
If /prompt-drift -help or /prompt-drift -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 and -GitRange (default: last 20 commits). Get confirmation.
Step 3 - Scan
& "<SKILL_DIR>/scripts/prompt-drift-scan.ps1" -ProjectDir "<path>" -GitRange 20
Step 4 - Classification
For each drift:
- Benign: cosmetic or clarification changes
- Monitor: minor change, verify outputs
- Significant: output behavior likely changed
- Critical: safety posture weakened or output constraint removed
Step 5 - Write report
File prompt-drift-report.md in current working directory:
- Summary - drifts by severity.
- Drift table - critical first. Per drift: file, commit, date, changed sections, before/after snippet, severity.
- Open questions.
Step 6 - Summarize
State report path, highlight critical and significant drifts.
Usage
/prompt-drift # interactive
/prompt-drift <dir> # scan project
/prompt-drift -help