I stopped fixing my prompt by hand. Open models fixed it while I napped

Description and use

An automated prompt‑improvement loop using open‑weight models and AcruxCore raised a bank message router’s accuracy from 37 % to 60 % on Gemma 3 4B and from 69 % to 74 % on Gemma 3 27B, eliminating manual edits.

Problem

Manually editing prompts for a bank message router was error‑prone and time‑consuming.

Outcome

The prompt‑improvement loop, driven by open‑weight models and executed via AcruxCore’s MCP, eliminated manual edits and improved intent‑routing accuracy from 37% to 60% on Gemma 3 4B and from 69% to 74% on Gemma 3 27B, at a cost of $0.38.

Workflow

• Create lazy prompt (v1) and store via MCP
• Load 154 training examples with intent definitions
• Run optimizer to draft rewrites and judge scores
• Promote winning rewrite as new prompt version (v3)
• Evaluate on 308 held‑out messages

Evidence

Fixing a prompt by hand is a loop I hate. I rewrite it, test it by hand, redeploy, and find that I fixed two cases and broke a third.

Source

Gemma 3 4B 37% → 60% on 308 held‑out messages

Source

The router never changes. AcruxCore stores each prompt as numbered versions, and an alias is a name that points at one version.

Source

The optimizer wrote those definitions from the 154 training examples. On held‑out messages, wrong labels fell from 45% to 40%.

Source