Tidy Up Your Account

Using the LLM Prover MCP tools, help me tidy up my account. SAFETY FIRST: this recipe is mostly read-only, and you must NEVER delete anything without my explicit, per-item approval. 1. Confirm the tools you need are available (list the tools). 2. Take inventory, read-only: list_rag_stores, list_rubrics, list_system_prompts, and (if I use benchmarks) list_benchmarks + list_benchmark_runs for the ones I care about. Also list_comparisons / list_evaluations if relevant. 3. Propose cleanup candidates WITH REASONS -- do not just dump the lists. Flag things like: stores stuck "dirty" or empty, duplicate or clearly-superseded rubrics / system prompts, old run records beyond what I need. For anything that is referenced by something else (a store or rubric used by an active benchmark), say so -- deleting it will affect that benchmark. 4. Delete ONLY what I approve, one item at a time with an explicit confirmation for each. Never batch-delete on a single yes; confirm each item. For a store or rubric that is in use by an active benchmark, deletion requires confirmed_name matching the name exactly -- tell me it is in use, that deleting it will freeze or affect that benchmark, and only proceed with the confirmed_name once I have said yes to THAT specific item. 5. After each deletion, confirm what was removed. At the end, summarise what was deleted and what was kept. NEVER delete speculatively, in bulk, or to "clean up proactively" -- only on my explicit per-item yes. If you are unsure whether something is safe to delete, leave it and tell me why. ON ANY FAILURE (a delete is rejected, an item is in use): report it honestly and move on; do not retry a destructive action in a different way to force it through.

Goal

A guided, safe clear-out of the clutter an account accumulates – stale stores, duplicate rubrics and system prompts, old runs – where the agent does the finding and the explaining, and you make every delete decision. Nothing is removed without your explicit say-so.

When to use

Reach for this when your Stores, Rubrics, System Prompts, or run history have built up and you want help deciding what to clear, without risking something you still need.

How it works

Read-first, delete-last. The agent inventories everything read-only and proposes candidates with reasons before anything is touched.
Per-item consent. Deletions happen one at a time, each with its own explicit confirmation – never a single blanket “yes, delete all”.
In-use awareness. A store or rubric referenced by an active benchmark requires confirmed_name to delete and will affect that benchmark; the agent flags this and only proceeds on a specific yes for that item.

Verification

The inventory step only listed; nothing was deleted before you approved it.
Each deletion had its own explicit confirmation (no bulk delete on one yes).
In-use items were flagged with their impact before deletion, and the final summary matches exactly what you approved.