Your AI roadmap, the index for everything that follows

This post is the syllabus for AI Foundations, a self-paced training course we publish as a blog series. Every entry below links to a published lesson. Bookmark this page. As new posts ship, they slot back into the right track here.

The order is opinionated. Track 1 is mandatory for everyone. Tracks 2 and 3 are for anyone who wants to use AI productively. Tracks 4, 5, 6 are the deeper paths for builders. Track 7 covers the future direction. Track 8 is provider-specific deep dives. Track 9 is specialised topics that fork off whenever they are relevant.

You do not have to finish the map. You do have to start at the top.

Track 1: orientation

Where you build the mental model. Skip nothing here. Each post is short.

• What AI actually is The no-jargon primer. AI vs ML vs LLM, training vs inference, parametric vs context knowledge, what an agent is.

• The AI vocabulary you'll keep hearing Working glossary of the terms that show up in every AI article. Skim it once, return to it as needed.

• Inside an LLM, how language models actually work Tokens, embeddings, attention, prediction. Non-mathematical. Explains why LLMs hallucinate, miscount, and forget.

Track 2: using AI well

The skills that turn "I have access to ChatGPT" into "I get useful work out of AI every day".

• Talking to AI well, prompt engineering for everyone Six patterns we use daily. State the role, show the format, ground in source, separate task from data, ask for reasoning, iterate in place.

• AI tools worth your time in 2026 The tools we keep in our daily flow. ChatGPT, Claude, Gemini, Perplexity, Cursor, Claude Code. What each is best at and where each falls down.

• AI in your daily workflow, for non-coders Eight concrete patterns for embedding AI into knowledge work. Drafting, summarising, voice notes, comparison, rubber-ducking.

• Pair-programming with AI, for working engineers Editor integrations, CLI agents, PR review on tap. When to trust, when to verify, when to write it yourself.

Track 3: knowing the limits

Pattern-matching when AI is the wrong tool. Strong AI users skip more often than they reach.

• What AI is bad at, and why Exact maths, counting, post-cutoff facts, single-correct-answer domains. Failure modes and underlying mechanisms.

• Hallucinations, why models invent facts The most famous AI failure mode, mechanically explained, with practical mitigations.

• Prompt injection, the trust boundary How untrusted input subverts AI systems. Layered defences that actually hold up.

Track 4: going deeper

Optional. Take this track if you want to make decisions about AI for a team, or build with it in your own apps.

• Embeddings and semantic search What an embedding is. Why "cat" and "kitten" live close in vector space. How to build search that knows what you meant.

• RAG, retrieval-augmented generation How to ground a model in your own data without fine-tuning. The full pipeline and where each stage breaks.

• Agents, tools, and the loop When an LLM gets to call functions, see results, and iterate. What changes and what breaks.

• Fine-tuning vs prompting vs RAG Three ways to specialise a model to your data. The decision framework, with cost and capability tradeoffs.

• Retraining, when and how to update a model When fine-tuning earns its operational cost, and the workflow when you do commit.

Track 5: building with AI

For coders shipping AI features into production.

• First call, integrating an AI API Production-shaped first integration. Auth, retries, timeouts, observability, the small things that separate prototype from production.

• Structured outputs, making the model return data Force the model to return JSON that fits your schema. The capability that lets AI fit into normal pipelines.

• Tool calling, exposing your code to the model Designing tools, handling the loop, keeping agents safe. The engineering side of agents.

• Cost, caching, and prompt engineering for production AI feature economics at scale. The levers that move the bill, in order of leverage.

• Evaluation harnesses The discipline that lets you ship AI safely. Eval sets, graders, when to run, what to do with failures.

• Production safety, guardrails, and the kill switch The defence-in-depth layers that turn an AI feature into a production system that does not become an incident.

• Harness engineering, beyond prompt engineering The discipline of designing the system the model lives in. Tools, context, instructions, environment, feedback.

• Context engineering, beyond prompt engineering Selection, compression, freshness, prioritisation. The skill that determines whether a long-running AI feature stays useful.

• Coherence, keeping AI consistent across long sessions Long-document drift, multi-turn contradictions, agent decision drift. Patterns that hold output coherent across length and complexity.

• Content moderation with AI Filtering inputs and outputs. Provider tools, custom classifiers, LLM-as-moderator, and the trade-offs.

Track 6: specialised topics

Read these when relevant to your work or context.

• AI ethics and bias, the practical version Where bias enters models, how to test for it, what mitigation looks like in code, and the limits.

• AI in African contexts What does and does not generalise from US-trained models to East African use cases. Swahili, code-switching, mobile-first deployment.

• Privacy and on-prem AI When you cannot use cloud APIs. Open-weight, self-hosting, the actual cost and capability gap.

• AI for non-technical leaders How to evaluate vendors, scope projects, set expectations, and not get sold the wrong tool.

Track 7: where AI is going

The trends, the protocols, and the things to try if you are starting out.

• Where AI is going, trends to watch Long-arc directions: long context, agents, multimodality, on-device, reasoning models, cost.

• MCP, the standard wiring AI tools to your data The Model Context Protocol that is becoming the universal AI integration layer.

• Open-weight models, the AI you can run yourself Llama, Mistral, DeepSeek, Qwen. When self-hosting genuinely makes sense.

• Cool AI things to try if you are just starting Hands-on weekend projects. None require coding. They build the intuition no amount of reading will.

Track 8: provider deep-dives

The frontier providers are not interchangeable. Each has features and ecosystems beyond the bare model. These posts cover what each provider does well and where each fits.

• Claude deep-dive, the features that set it apart Skills, schedules, sub-agents, MCP, projects, computer use, artifacts. The Claude ecosystem feature-by-feature.

• OpenAI deep-dive, beyond GPT Custom GPTs, the GPT Store, Assistants, Realtime, Code Interpreter, image and video. The full OpenAI surface.

• Gemini deep-dive, AI in your Workspace Google Workspace integration, search grounding, multi-modal default, long context, Vertex AI. Where Gemini wins and where it does not.

• GitHub Copilot deep-dive The veteran coding assistant. Autocomplete, Chat, Edits, Workspace, Code Review, Spaces. What has aged well and what has not.

• The wider field of coding agents Beyond the big three. OpenCode, Aider, Continue, Cody, Codex CLI, Windsurf, Devin, Tabnine. Where each fits.

How to use this map

A few rules of the road:

1. Track 1 is mandatory. It is short. Most "AI is overhyped" takes come from people who skipped it, used a model badly, and concluded the tools are bad. The mental model is non-negotiable.

2. Track 2 is the highest-leverage on the entire map. Two hours of deliberate practice with the prompt engineering post will save you hundreds of hours over the next year. We are not exaggerating.

3. Stop when it stops being useful. A product manager probably stops at Track 2 plus the parts of Track 3 that interest them. A staff engineer probably wants through Track 5. There is no prize for finishing.

4. Come back to this page. When new posts ship in the series, they index here. The series grows over time; the map stays oriented.

5. Suggest topics. If you want a deep dive on something specific (a model, a tool, a use case), reach out via the contact form. We write the next post based on what enough people are stuck on.

Track 1 takes about 90 minutes. The prompt engineering toolkit in Track 2 takes another hour. After two and a half hours you are out of "I do not know what I am looking at" and into "I can use this tool deliberately".

That is what the rest of this series is for.