Prompt Engineering in Production Is a Systems Problem
The prompt is the smallest part of the system. What determines output quality is the context, the constraints, the feedback loop and the evaluation set.
Tutorials
Build guides that include the config that broke
Step-by-step for agents, retrieval pipelines, workflow automation and GEO infrastructure.
Every guide here was written against a system that is running in production, handling real tenants and real traffic. When a step depends on a version, the version is in the text, not implied.
I write down the part that broke. Most tutorials skip it — they show the happy path and leave you stuck at the first failed tool call. If something failed for me, it is in the post with the error message and what fixed it.
Start here
How to Secure an AI Agent That Can Call Tools
Harden an agent that reads the web, calls APIs and writes files — where injected instructions land, plus least privilege, approval gates and audit logs.
Tutorials
Write GEO Content That Gets Cited by Answer Engines
Structure pages so generative engines can lift them — a 40-60 word direct answer, extractable facts, matching structured data, and crawler access.
Build Your First RAG Pipeline That Doesn't Hallucinate
A working retrieval-augmented pipeline with chunking, reranking and an evaluation harness — plus the three failure modes that produce confident wrong answers.
How to Evaluate an LLM Pipeline So You Can Change It Safely
Build an evaluation set from real production traffic, assert on properties instead of exact strings, and wire it into CI so prompt changes ship on evidence.
How to Cut Your LLM API Cost Without Breaking Quality
Measure first, then route, cache, shorten, batch, and quantize — a practical order of operations for cutting LLM spend without shipping worse output.
Tutorials questions
What does these tutorials assume about your setup?
A single Linux box, a provider API key, and Python 3.11 or newer. Most guides deploy one service and wire it to one model endpoint. Anything heavier is called out explicitly at the top.
Do I need a vector database to follow along?
No. Several guides run entirely on a local embedding index so you can get the pipeline shape right first. Where a hosted store genuinely helps, the guide says which one and why.
Are the version numbers kept current?
I re-run the guides when a major version lands and update the post date at the top. If you hit a version that has moved on, the contact page is the fastest way to get it fixed.
Keep going
- Comparisons Choosing between two tools is a different job from wiring up one.
- Reviews Long-term notes on what actually held up after months in production.