Skip to Content
agent-stats

Review recurring prompts

Identify repeated corrections and decide what to document in project instructions.

Run a review across all projects for the last seven days:

agent-stats prompts --all --since 7d

You can also open Prompts in the dashboard, or press m in the interactive terminal until Prompts is selected.

Read the review

  • Recurring groups matching flagged prompts found in at least two sessions. Check whether an instruction in AGENTS.md or CLAUDE.md would save you repeating the request.
  • Prompt signals by model compares flagged and total prompt counts by the preceding model and provider when that link is available. Treat the counts as signals and compare them with the number of prompts each model handled.
  • By folder helps you find where flagged prompts are concentrated.
  • Flagged prompts let you inspect the excerpt, source session, and signals behind an entry.

Signals include repetition, negation, profanity, and all-caps text. Pasted logs or quoted text can trigger them. Open the source session before acting on a flag.

Export or filter

agent-stats prompts --project ~/work/app --since 30d
agent-stats prompts --all --since all --json

The command defaults to seven days. Use --since all for the full index. JSON from this command can contain prompt excerpts, so review it before sharing.

Use agent-stats query prompts for analysis without prompt excerpts by default. Open a selected full prompt with query prompt only when its wording is needed.

On this page