ai-talk-show/prompts/producer_brief_007.md

Producer Brief — Episode 007: Context Engineering (It's Not Just Prompts)

Topic

The discipline of getting the right information to the model at the right time — and the 2026 reframing that "prompt engineering" was the wrong name. Context engineering splits into four levers: instructions (system prompt / behavioral framing), retrieval (RAG / grounded search), memory (short-term conversation + long-term persistent state), and tools (the function-calling surface, increasingly standardized via MCP).

Why It Matters

This is the connective tissue for everything in episodes 004-006. MoA is about context from peers; verification is context about correctness; computer use is context about the screen. Context engineering is the meta-skill: a coding agent that forgets your architecture decisions after two sessions is broken even if its prompt is perfect. The high-signal point — most agents fail not on model quality but on context design (what's in the window, what's retrieved, what's remembered, what's scoped out).

Seed Question (goes to Leonard first)

"People keep calling it 'prompt engineering' like it's about writing a clever instruction. But the 2026 take is that the real discipline is context engineering — four levers: instructions, retrieval, memory, and tools. A coding agent with a perfect prompt that forgets your architecture after two sessions is still broken. So what actually separates an agent that has context from one that's just got a big prompt?"

Context Notes (background depth — NOT a script)

• Four-component model (Sourcegraph, May 2026; Anthropic guidance): instructions, retrieval (RAG/grounded search), memory (short + long-term persistent), tools (MCP-standardized function surface).
• Mem0 / Supermemory framing: shift from crafting better prompts → designing memory architectures; separates functional prototypes from agents that provide genuine value.
• Anthropic "Effective context engineering": keep context informative yet tight; dynamically retrieve context at runtime rather than stuffing the window.
• Long-term memory is the unsolved part: external scratchpads/files/memory stores that survive beyond a single window and are recalled deliberately.
• Builder angle: Hermes itself — persistent memory across sessions, MCP tools, retrieval. Is the talk-show's own Leonard/Charlie design (no persistent memory per their CORE RULES) a feature or a bug for a show? Good tension to land on.