AI & LLM Testing Learning Path: Start Here

In short: you learn AI testing in 6 milestones by building one evaluation harness for a support bot and growing it — golden answers, attack prompts, structured-output checks, RAG and agent tests. It runs without an API key (the "model" is a stand-in), so you can swap in a real model whenever you like.

How to use this page

Read this page once to see the plan. Then, for each milestone, follow the same four steps: learn → build → check → commit. Come back here whenever you're unsure what to do next.

Before you start

• Python 3.10+ (for pytest); Node.js 20+ from Milestone 3 (for promptfoo).
• Basic Python: functions, dicts, assert.
• The lab — evals/bot.py and friends.
• Optional, for real-model milestones: an API key for Anthropic or OpenAI.
• Helpful: the API testing cheat sheet if you'll call a model's API.

The pages in this learning path

Page	What it's for	When to open it
This roadmap	The plan and self-checks	At the start of each milestone
AI/LLM testing cheat sheet	Learn each idea, with runnable code	The "learn" step
Milestones & Mini-Projects	Tasks, expected results, solutions	The "build" and "check" steps
Quick Reference	Assertions, OWASP LLM Top 10, metrics, attacks, tools	Any time
Best Practices	Habits that make evals trustworthy	After Milestone 2, then before shipping
AI product & LLM testing guide	The service and AI-augmented QA	For the deeper "why"

The milestones

Milestone	You learn	You build	Rough time
1	Why AI differs, golden datasets, the harness	Accuracy tests	1 week
2	Hallucination, prompt injection, OWASP LLM Top 10	An attack set	1–2 weeks
3	Structured output, determinism, promptfoo	Config-based evals	1 week
4	RAG: retrieval, groundedness, citations	RAG tests	1–2 weeks
5	Agents, tool limits, LLM-as-judge	Agent + judge tests	1–2 weeks
6	Regression, cost, an evaluation report	A CI eval and report	1 week

Times assume about 5 hours a week.

For each milestone: learn (cheat-sheet sections) → build (the tasks) → check (run the tests / solution) → commit.

Milestone 1: The harness

Learn: Why AI testing is different · What to test · Evaluation, not pass/fail · Golden datasets · Metrics · The harness

Build: Accuracy tests

Check yourself:

• Why can't you assert output == expected for an LLM?
• Why store facts instead of exact strings in the golden set?
• What does the must_not_include list protect against?

Milestone 2: Attacks & safety

Learn: Accuracy & hallucination · Prompt injection & the OWASP LLM Top 10 · Quick Reference: Attack prompts, OWASP LLM Top 10

Build: An attack set

Then read: Best Practices, sections 1–4.

Check yourself:

• What two things must you check for each attack prompt?
• What is indirect prompt injection?
• Why is injection resistance a 100% threshold, not 95%?

Milestone 3: Output & promptfoo

Learn: Structured output · Determinism & temperature · promptfoo · Quick Reference: promptfoo

Build: Config-based evals

Check yourself:

• How do you check JSON output properly?
• Why run each case several times with a real model?
• When would you choose promptfoo over pytest?

Milestone 4: RAG

Learn: RAG testing · Quick Reference: RAG

Build: RAG tests

Check yourself:

• Why test retrieval and the answer separately?
• What is groundedness?
• What should a RAG bot do when no document matches?

Milestone 5: Agents & judges

Learn: Agents & tool use · LLM-as-a-judge · Quick Reference: Agents, Judge rubric

Build: Agent + judge tests

Then read: Best Practices, sections 6–7.

Check yourself:

• Why enforce agent limits in code, not the prompt?
• How do you calibrate an LLM judge?
• When must a human review, not a judge?

Milestone 6: Regression, cost & report

Learn: Regression after changes · Cost & latency · Bias & safety · AI-augmented QA

Build: A CI eval and report

Then read: Best Practices, sections 5, 8–11.

Check yourself:

• Why pin the model version?
• How can a new model improve overall but still be a regression?
• What belongs in an evaluation report?

When you get stuck

Problem	What to do
Tests pass but feel too easy	The stand-in bot is deterministic and careful by design — swap in a real model to see failures
A real model's answers vary	Test facts/meaning, not exact text; run each case several times
No API key	Everything here runs on the stand-in; add a key only for the real-model tasks
promptfoo can't find the provider	Use python:provider.py with the path relative to the config file
Judge disagrees with you	Tighten the rubric; calibrate on more human-graded samples

What's next

• Security of AI features: security cheat sheet and OWASP LLM Top 10.
• Load-test AI endpoints: performance cheat sheet.
• Build with models: the Claude cheat sheet and MCP & AI agents guide.
• Certifications: ISTQB CT-AI (AI Testing) and CT-GenAI (Testing with Generative AI).

Good resources: OWASP Top 10 for LLM Applications, the promptfoo docs, and Anthropic's and OpenAI's own guides on evaluating models.