AI Native Playbook — Testing skill for Claude Code

[🇧🇷 Português (Brasil)](README.pt-BR.md) · 🇬🇧 **English**

AI-Native Engineering Playbook

**How I run coding agents on regulated production codebases without losing control.**

This is the operating system I use to ship software with AI coding agents (Claude Code, Codex, Copilot) on systems where mistakes cost money: an international payments platform (card-network messaging, real-time settlement, FX, AML reporting) and [Sinal](https://sinal.andrecristino.cloud), a multi-tenant AI-agent SaaS that I designed, build and operate. Every rule here exists because something went wrong without it.

**What is real and what is not.** The rules, the incidents and their dates are real. Client and partner names are absent on purpose, and so is anything operational: hostnames, account identifiers, volumes, totals. The worked example is fictional. Where a figure appears it is a ratio, or a count of my own artefacts. Never a client's business number.

Think of it as a harness. The contracts, runbooks, specs, memory and evidence discipline that let an agent do real work on a codebase without turning it into a liability. Prompts are the smallest part of it.

Two chapters go further than the harness, because there the agents are also the product. [Evaluating an LLM product](docs/09-evaluating-llm-products.md) covers the evaluation harness I built into a live agent SaaS: persona stress tests, LLM-as-judge against a rubric, runs keyed by the hash of the prompt so comparisons stay honest, hallucination scored apart from an honest gap. [Security hardening](docs/08-security-hardening.md) covers the seven-front audit method and the toolkit that *proves* a fix instead of assuming it. If you only read one of these ten documents, read [07. Evidence discipline](docs/07-evidence.md). It is the one that changes how you work tomorrow.

The model in one picture

            ┌──────────────────────────────────────────────────────────────┐
            │                       WORKING AGR