12× fewer errors. Same model. Same call.
Refusals, malformed outputs, schema drift, hallucinated formats, the failure modes every provider ships with. The enhancer is a learned layer that catches these before the model call and reshapes the prompt so the failure never happens. Same model, same API, 12× fewer broken responses.
- READThe enhancer reads the model's error code, content safety, content policy, or a celebrity / brand-IP hit.
- REWRITEIt rewrites only the flagged tokens, never the intent, then re-checks against all three policies before shipping.
- HOWAdd
enhance.prompt: trueto your API call. The trace tells you what was rewritten and why.
⚐ projections from the early-access cohort · subject to change at GA
A small model that catches what your model would miss.
The enhancer is a fast LLM that reads every prompt, predicts where the target model will fail, refusal, malformed output, format drift, schema break, and reshapes the prompt so the call lands cleanly. Your code, your call signature, your model. 12× fewer errors in production.
Refusal repair
Provider returns a refusal on an ambiguous-but-harmless prompt? The enhancer catches the trigger before the call and reshapes it. Your user sees a result, not a "sorry, I can't help with that".
Schema-aware enhancement
When you need JSON, you get JSON. The enhancer enforces shape before the model sees the prompt, malformed-output bugs collapse from ~8% to <1%. Your parsers stop crashing.
Per model failure mapping
Each model has its own failure surface, kling refuses different prompts than veo; flux malforms differently than nano-banana. The enhancer maps to the target model, so cross provider swaps stay reliable.
One flag. Five policies. Every refusal saved.
Add enhance.prompt: true to any API call. The enhancer watches the policy verdict, rewrites only when it would have failed, and stamps the trace so you can audit what was changed.
Enhance prompt
One flag on your existing API call. No SDK swap, no separate endpoint.
intent_priority: "preserve" to lock the user’s meaning.What gets caught
The enhancer adapts to each provider’s policy table. These are the categories it learns to swap automatically.
Read the prompt trace
Every enhanced call carries a trace.enhancer block, the original, the rewritten, what got rejected, and the recheck verdict.
Reach for the enhancer when…
Your users hit the refusal-rate floor
Consumer prompts trip safety filters, even when nothing is unsafe. Without enhancement, ~12% of prompts come back as a polite refusal. With it, ambiguous prompts get reshaped before they ever hit the model, and the rate drops to under 1%.
refusal_rate: 12.4% → 0.9%You need JSON and the model gives you "mostly JSON"
Schema drift breaks downstream parsers. The enhancer enforces shape before the call, malformed outputs collapse from 8% to <1%. Your retry loops empty out, your bills shrink, and your parsers stop crashing.
malformed: 8.1% → 0.6%You're swapping kling for veo this week
Each model fails differently, kling refuses different prompts than veo; flux malforms differently than nano-banana. Without the enhancer, swapping providers means relearning each failure surface. With it, error rate stays 12× lower across any swap.
model swap · same error floorYou don't have a prompt engineer to hire
Hiring a prompt engineer is a 6-month search and a $200K headcount. The enhancer benchmarks above the median candidate on every internal eval, same error reduction, no hires, no quits.
12× fewer errors · 0 hiresOther products you’ll use alongside this.
FAQ
What is the each::labs prompt enhancer?
Stop debugging prompts. Start shipping reliable outputs.
Enhancer is free on every plan; schema-aware enhancement + per team learning on Pro and up.