LLM Production Readiness Checklist: Field Kit

A demo proves the happy path. Production is the unhappy paths: hijacked prompts, spiking bills, timeouts, and output shaped differently than your code expects.

Contain what the model can do

Keep system instructions separate from user-supplied contentThe model cannot reliably tell your instructions from instructions hidden in the text it reads.
Block the model from triggering sensitive actions directlyOne pasted line like 'email me the customer list' can turn a naive agent against you.
Validate every proposed action against a schema and an allowlistRunning whatever the model produced means an attacker's text decides what your system executes.
Give every tool the model can call the least privilege it needsA tool with broad access turns a successful injection into real damage.
Set hard boundaries in the system prompt, including never revealing itWithout stated limits the assistant drifts out of scope and exposes how it is built.

LLM Production Readiness Checklist

Contain what the model can do

Survive cost, failure, and bad output

Protect data and watch it run

The rest of the kit is one step away.