The LLM is a stateless box. Your job isn't to write prompts — it's to design the perfect context window every single time.
Every time an agent runs, it starts from zero. Context Engineering is the art of assembling exactly the right information — no more, no less — so the agent can succeed.
Every agent prompt is assembled from these 6 blocks. Some agents need all 6, simple ones need just 2-3. The key is knowing which blocks to include and filling them precisely.
Who is this agent? What broad outcome should it achieve? Define the role, scope boundaries, and success criteria — but don't micromanage the steps.
Every API, MCP tool, and webhook the agent can call. Specify exact input/output schemas — the agent must know what's available and how to use it.
What private data does the agent need? Notion databases, files, external sources. Use intelligent retrieval — let the agent ask for data, don't dump everything.
Outcome-defined, not prescriptive. Describe what each phase should achieve, not step-by-step if-then logic. Let the model choose the path.
How does this agent handle state? Session persistence, compaction rules (what gets summarized vs. dropped), and key-value stores for bulky intermediate data.
Should this be one agent or many? If two phases need completely different context windows, split them. Define handoff protocols between sub-agents.
Moving from "prompt engineering" to Context Engineering changes how every skill and agent gets built.
From idea to running agent — every project follows this flow through your Notion system.
Every database serves a specific role in feeding context to your agents.
Business context: who you are, what businesses you run, brand voice, service packages. Feeds Block 1 (Identity).
Project intelligence: every tool, agent, and automation tracked with PRDs containing all 6 CE blocks. Feeds Blocks 2-6.
Reusable prompts, templates, agent personas. Single source of truth for all system prompts. Feeds Block 1.
Session continuity: what happened, what blocked, next steps. Cross-session memory. Feeds Block 5 (Memory).
Learning state: curriculum levels, skills mastered, what to learn next. Context for skill-building decisions.
Priority actions for the next session.