Resources

Where to read next.

A short, checked list rather than a long one. Everything here is free to read, and each entry says who wrote it — because documentation from the company that built a model and a blog post about that model are not the same kind of evidence.

What the labels mean

  • Official — written and maintained by the people who build the thing. Correct by default; check the date, not the claim.
  • Community — collected by volunteers. Broad and genuinely useful, but nobody guarantees any single entry is current.

If a page here ever cites a company's own benchmark of its own product, it will say so plainly.

Start with the documentation

Nearly every question a beginner has is answered properly in one of these four, and they stay current in a way that tutorials do not.

The concepts worth learning

These are the terms you will meet everywhere once you move past basic prompting. One plain sentence each, and which phase covers it — so a word you don't know never stops you.

Context engineering

Choosing what goes into the model's limited window, in what order, and what gets left out. The skill this whole book is organised around.

Phase 2

Memory architectures

Files or databases that let an assistant pick up where it left off, instead of starting from nothing every session.

Phase 2

Conversation summarisation

Replacing a long history with a short record of what was decided, so the important parts survive and the bulk doesn't.

Phase 2

Prompt compression

Shrinking text before it is sent, so more of what matters fits in the same space for the same money.

Phase 2

Long-context management

Deciding what to keep and what to drop first once the window is full. Every long session eventually runs into this.

Phase 2

Retrieval-Augmented Generation (RAG)

Looking up the handful of documents relevant to a question and pasting them in before the model answers, rather than hoping it already knows.

Phase 2

Knowledge graphs

Storing facts as connected things — people, projects, dates — instead of loose paragraphs, so relationships can be looked up directly.

Phase 3

Prompt optimisation

Improving a prompt by measuring versions against a test set, rather than rewriting it until it feels better.

Phase 1

Tool calling (function calling)

Letting the model ask your code to do something — search a database, send a file — and then use the result.

Phase 4

MCP servers

A shared standard for exposing your tools and data to any assistant, so you write the connection once instead of once per product.

Phase 4

Agentic workflows

The model runs in a loop, choosing its own next action and reacting to what happened, instead of answering once and stopping.

Phase 4

Planning versus execution

Separating "work out what to do" from "do it". Keeping the two apart makes failures much easier to find.

Phase 4

Reflection and self-critique

Having the model check its own work against the goal before it hands anything over.

Phase 4

Multi-agent systems and A2A

Several assistants with different jobs passing work between them, and the conventions they use to talk.

Phase 4

Tool orchestration

Running several tools in a reliable order, and deciding what happens when one of them fails.

Phase 4

Local models (Ollama, vLLM)

Running a model on your own machine or server — slower and smaller, but private and free per token.

Phase 3

Evaluation frameworks

Measuring whether a change actually improved anything. Without this, every other technique here is guesswork.

Phase 3

The Markdown files people keep

Since an assistant starts each session with no memory, the common fix is to write things down where it can read them again. Most projects settle on a handful of plain Markdown files at the repository root.

PROJECT.md holds what the project is and why. MEMORY.md keeps facts and decisions worth surviving. SUMMARY.md is a short record of a long conversation, loaded instead of replaying the whole thing. TASKS.md tracks what is done, what is next, and what is blocked.

Alongside them sits a skills folder: one file per capability, describing how a job should be done. The assistant reads only the titles until one matches, then opens that file. That is progressive disclosure, and it is why the red band in the homepage demo disappears.

Is this an official standard?

No. These filenames are a widely copied convention, not a specification, and different tools expect different names — some want CLAUDE.md, some read a memory-bank/ folder, some use none of this. Pick one scheme and keep it consistent inside a project; that matters far more than which names you chose.

a common layout

project/
├── README.md      for humans
├── PROJECT.md     what this is, and why
├── MEMORY.md      decisions worth keeping
├── SUMMARY.md     the long conversation, shortened
├── TASKS.md       done · next · blocked
├── CHANGELOG.md   what changed, when
├── prompts/
├── skills/
│   ├── coding.skill
│   ├── research.skill
│   └── writing.skill
└── docs/

Know something that belongs here?

This list stays short on purpose, but it should not stay wrong. Suggest an addition, or tell us when a link rots.