Live answer, requirement 2 of 12

Prompt engineering

I write the output shape first, version the prompt, budget its tokens, and compare model replies before shipping.

The requirementThe role requires hands-on prompt engineering, with LLM, GenAI and RAG experience behind it.

Try it

The demo sends your task to the same model twice, once as the bare question and once as an engineered prompt. That prompt carries a role, rules and a required JSON shape, and the reply is validated against that shape. A failed check goes back to the model once for a repair before you see the result.

How it works

Write the output shapeI fix the required keys and the allowed values before I write any wording, and the prompt states that shape verbatim. I use no provider schema mode, so the caller parses the reply strictly and returns nothing when a key is missing.
Version every templateI keep every template in a versioned store with named double-brace placeholders, and the render step reports which values are still missing. Each edit bumps the version number in the same write that changes the text, so we can roll back one prompt.
Compose per personaIn a layout editor I order 13 prompt sections into one system prompt for each persona. Only the system, chat and voice channels are augmented at runtime. Each live turn resolves the composed string in process, so no extra model call is spent.
Budget the contextI run an estimator that reports characters and estimated tokens per section, then fits the total against every model in the catalog. Retrieved context stays inside my retrieval defaults, top-k 5 and 6000 characters, so documents cannot crowd out the instructions.

Tools

Tool	What it does here	Status
Prompt composer, estimator and simulator	Orders the sections into one system prompt per persona, reports token cost against each model window, and runs the prompt against a model.	Available
Versioned prompt template store	Holds templates with named variables and bumps a version on every update, reachable today only through the internal tool surface.	Early access
Multi-provider model gateway	Routes a prompt to a provider with key handling, rate limits and per-call logging, and sets prompt-cache breakpoints on the Anthropic adapter.	Available
LLM-judge evaluation harness	Scores model answers over a versioned dataset on named rubric dimensions, with hard-fail flags and a pass or fail exit code.	Available
Public prompting demo	Runs the side-by-side comparison on this page on Cloudflare Workers AI, with Llama models, a schema check and one repair retry.	Demo only

Available means shipped and in use. Early access means built but not generally available. Roadmap means designed, not built. Demo only means it powers this public demo, not customer deployments.

The full write-up covers access control, evaluation and how this runs inside a customer environment.

Read it on gusit.de

The other answers

LLM applications in productionEvery model call goes through one gateway, guardrails run on both sides, and each task gets a measured model choice.Retrieval-augmented generationGrounded, cited answers over the customer's own documents, with access rules applied before retrieval.AI agents and orchestrationI build agents that call your systems as registered tools, and the workflows that trigger and sequence them.Request classification and routingI build automation that classifies each inbound request, routes it by role, and holds risky changes for human approval.AI workflow orchestrationI design the steps, contracts, and data flows, then ship them as running workflows on customer systems.API and integration designI publish the API contract first, gate it with keys and scopes, then build integrations against it.Hybrid enterprise integrationI connect AI services to your identity, applications and data, then run them where your policy allows.Prototype to productionI take a prototype through generated test suites and merge review gates to a zero-downtime deploy that rolls back on failure.Approvals and exception handlingI put a person between the agent and the change, with gated tools, a pending queue, expiry, and a record.Evals and output validationI score answers against a versioned rubric, trace failing calls to the integration that broke, and tune latency per call.Client-facing deliveryI turn engineering work into updates that client architects, security leads and business owners can each act on.

Gus IT LLC. Contact gus@gusit.de. Claude is a product of Anthropic PBC. Cloudflare and Workers AI are trademarks of Cloudflare, Inc. Llama is a trademark of Meta Platforms, Inc. Other names are trademarks of their respective owners. Gus IT LLC is an independent engineering firm and is not affiliated with or endorsed by these companies.