- 1The context windowEverything the model reads in one call, in order; the middle counts least, and a full window rots.
- 2Priming and sycophancySix things decide what a line in the window is worth; the first lines set the tone, and a push with no new fact bends it.
- 3Tools and the agent loopThe model writes a call, a program runs it and calls the model again; MCP brings the tools to every AI app.
- 4Subagents and orchestrationA task in a window of its own; code or the model picks the next step; checks around the loop.
- 5Prompt injectionThe model reads an order and a poisoned page the same way; six ways it arrives, and the lethal trifecta.
- 6What works against injectionWords in the prompt help a little; less reach, a person's yes and nowhere to send data hold.
- 7Evals and the golden setYour own questions with their right answers, five ways an answer lands, and a judge before it goes out.
- 8The kinds of memoryIdentity, procedures, knowledge, events and data: semantic, procedural and episodic memory, and the environment.
- 9instibrain in these wordsContext window, retrieval, agentic, structure: where each word of the series sits in one brain.
What people build around a model, in nine short lessons.
A model alone reads text and writes text. How a model works opens it. This series is what people build around it: the context window it reads, the tools it calls, the attack it meets, the test that says it works, and the kinds of memory a company gives it. Retrieval and cost have series of their own, and the lessons link them.
No code and no mathematics. Every Try it is words to type in a Claude window.
Nine lessons, from the context window to the memory.
Lessons 1 and 2 are the context window and what a line in it is worth. Lessons 3 and 4 are tools and agents, lessons 5 and 6 the attack on them and what stops it. Lesson 7 says how to know it works, lesson 8 names the kinds of memory, and lesson 9 puts instibrain in these words.
The map: every word, in its place.
The context window is what the model reads. Tools let it act, checks keep it safe, a test says it works, and the kinds of memory say what goes in. Retrieval is in RAG; what a token costs is in AI hygiene.
| Where | The words | Lesson |
|---|---|---|
| The context window | context, lost in the middle, context rot, compaction, context engineering, progressive disclosure | 1 |
| A line's worth | instruction hierarchy, hard and soft constraint, priming, sentiment, sycophancy | 2 |
| Tools | tool use, function calling, tool result, agent loop, MCP | 3 |
| Orchestration | subagent, workflow, agentic, guardrails, sandbox, human in the loop | 4 |
| The attack | prompt injection, jailbreak, tool poisoning, memory poisoning, exfiltration, lethal trifecta | 5 |
| The defence | least privilege, allowlist, taint tracking, defence in depth, red teaming | 6 |
| The test | benchmark, golden set, abstention, hallucination, grounding, regression, live judge | 7 |
| Memory | working memory, semantic memory, procedural memory, episodic memory, environment | 8 |
| The stack | context window, retrieval, agentic search, structure, curation, long context | 9 |
Running retrieval beside your brain: what to index, the job, the cuts, hybrid search, the reranker, the prompt, the measure, and where the pages sit.
why it mattersContext engineeringWhat a brain keeps in the context window, and what stays out.
every seriesGuidesThe four kinds, what you need, and how a lesson works.