Skip to main content

curompt

curompt is a CLI for analyzing, evaluating, and optimizing LLM prompts. It helps teams measure quality (schema-fit, self-consistency), detect regressions, and produce reproducible prompt reports directly from a terminal or CI workflow.

Why curompt?​

  • Local-first & reproducible – runs without sending prompts to third-party services unless a provider is explicitly selected.
  • Contract-first – validates prompts against JSON Schema / OpenAPI definitions to keep expectations explicit.
  • Scoring & suggestions – generates 0–100 health scores with actionable recommendations (token trimming, rule separation, few-shot cleanup).
  • Report automation – produces Markdown/HTML summaries that slot into pull requests and runbooks.

Core capabilities​

AreaHighlights
Static analysisSection heuristics, forbidden rules, duplicate detection, token budget
Dynamic evaluationMulti-sample schema fit, self-consistency, latency, cost (per provider)
ScoringOverall + sub-metrics with history persistence (SQLite)
SuggestionsAutomatic wording, structure, and formatting improvements
ReportingRich TUI output or file export via --output / --single-output

Typical workflow​

make build
./bin/curompt scan --path prompts/ --output reports
./bin/curompt eval --file prompts/onboarding.md --provider claude
./bin/curompt suggest --file prompts/onboarding.md > suggestions.md
  1. Scan a repository or prompt folder to generate baseline metrics.
  2. Evaluate critical prompts with an LLM provider (Claude, OpenAI, Gemini, local).
  3. Suggest improvements and track them in git.

Scan note: curompt scan uses prompts already saved in the local DB. When no history exists for the requested path, the CLI offers to auto-collect from Claude Code or Codex logs (Cursor support lands in v1.1).

Next steps​