Intelligence

Prompt Drift

Catch silently degrading prompts

npx promptreports-cli prompts audit

Walks lib/ and app/api/ to find LLM call sites. Checks git history for modifications. Flags deprecated models.

prompts audit is a static analyzer for your LLM integrations. It finds every place your code calls Anthropic/OpenAI/OpenRouter (system prompts, user prompts, tool definitions), cross-references with git history to flag recently changed prompts (highest drift risk), identifies deprecated models (gpt-3.5-turbo, claude-3-opus, claude-3-5-sonnet-old), and flags oversized system prompts (>800 tokens).

Prerequisites

lib/ and app/api/ directories
Git repo

Flags & Options

Flag	Description	Default
--days N	Recent-change lookback	7
--json	JSON output	—

Examples

Audit everything

npx promptreports-cli prompts audit

Every LLM call in your codebase.

30-day drift

npx promptreports-cli prompts audit --days 30

What prompts changed last month?

Output

Total call site count + recently changed (with git hash + commit subject) + model mismatches + token waste (oversized prompts).

  18 LLM call sites found across 5 files

  CHANGED RECENTLY (last 7 days):
  lib/ai/chat.ts:42
    a3b4c5d2 "fix: update claude-3.5-sonnet model"
    Model: claude-3-5-sonnet-20241022 → claude-sonnet-4-6
  lib/ai/generate.ts:89
    c5d6e7f8 "feat: add streaming support"
    System prompt: +240 tokens

  MODEL MISMATCHES (deprecated or outdated):
  lib/ai/legacy.ts:12  — gpt-3.5-turbo (deprecated Sep 2025)
  lib/ai/fallback.ts:23 — claude-3-opus (replaced by claude-opus-4)

  TOKEN WASTE (prompts > 800 tokens):
  lib/ai/analyze.ts:78 — ~1,200 tokens (consider trimming)
  lib/ai/summarize.ts:34 — ~950 tokens

What it reads and writes

Reads

lib/**/*.ts
lib/**/*.tsx
app/api/**/*.ts
app/api/**/*.tsx
git log

Writes

Nothing (read-only)

Free vs Pro usage

Free tier

Find deprecated models before they break
Spot prompt drift during code review
Document all LLM call sites for security audit

Pro tier

Upgrade

Auto-open PRs to update deprecated models
Prompt quality scoring — test every prompt against golden examples
Prompt version control with rollback
A/B test prompt variants in production

Pro tips

Run before every major deploy
Pair with `costs --by model` to find models to swap
System prompts > 800 tokens are a smell — extract to tools or skills

Free to try

Install the whole toolkit.

One command unlocks all 23 tools.

Start Free Back to all tools