Research Synthesis: Designing a Context Engineering Webinar

Date: May 2026 Context: Foundation research before designing a 60–90 min webinar on context engineering for a mixed-experience audience (general curious + intermediate LLM users).

This document synthesizes findings from four parallel research streams. Full detail in:

• 01 — Definitions and framings
• 02 — Resources bibliography
• 03 — Similar workshops
• 04 — Context vs prompt engineering

---

TL;DR

1. “Context engineering” is a real, ~11-month-old term-of-art with a clear origin story (Lütke → Chase → Karpathy → Anthropic, June–Sept 2025) and broad consensus on its meaning: the practice of curating everything the model sees at inference time, not just the prompt string.

2. The teaching landscape splits cleanly in two: consumer-tier (Andrew Ng, Wharton, IBM/Coursera) that stops at prompt patterns, and engineer-tier (Maven cohorts, AI Engineer Summit, LangChain webinars) that assumes Python and agents. There is no good 60–90 min webinar serving curious generalists + intermediate users with real context engineering content. This is the opening.

3. Two teaching frameworks have won. Use them as the workshop’s structural backbone:
   • Lance Martin’s four strategies: Write / Select / Compress / Isolate
   • Drew Breunig’s four failure modes: Poisoning / Distraction / Confusion / Clash
4. The non-engineer angle is undersupplied. Almost everything published assumes you’re building an agent. Almost nothing teaches the same mental model applied to everyday ChatGPT/Claude workflows (research, writing, analysis, document Q&A). That’s our differentiation.

---

What is context engineering (the workable definition)

The cleanest definition for a mixed audience, combining Tobi Lütke and Phil Schmid:

Context engineering is the practice of providing the right information and tools, in the right format, at the right time, so the LLM has everything it needs to solve the task — and nothing that gets in the way.

Karpathy’s OS metaphor (via Lance Martin) is the most teachable visual:

The LLM is like a CPU. The context window is like RAM. Context engineering is loading the right things into RAM at each step — and clearing out what’s no longer useful.

Anthropic’s “attention budget” frame is the best one-liner for why it matters: context is a finite resource with diminishing returns. More context isn’t better; better-curated context is better.

Why this is more than prompt engineering

The shift the field made between 2023 and 2025:

	Prompt engineering (2022–2023)	Context engineering (2024–2026)
Unit of work	The prompt	The trajectory
Window	4K–8K tokens	200K–1M+ tokens
Authored	Once, by hand	Dynamically, every turn
Ingredients	Instructions + a clever trick	Instructions + examples + retrieved knowledge + tools + memory + state
Failures	Bad wording	Poisoning, distraction, confusion, clash
Goal	Coax a chatbot	Engineer a system

Travel-agent illustration (the canonical side-by-side):

• 2023: “You’re an expert travel agent. Don’t exceed €200/night. Book Paris hotel.”
• 2026: System dynamically assembles: user profile from memory + the relevant clauses from the corporate travel policy via RAG + conference dates from calendar tool + live hotel availability from search tool. System prompt is small (~20%); the bulk is just-in-time context.

The competitive landscape (and the gap)

29 mapped offerings across Maven, DeepLearning.AI, Coursera, Anthropic Skilljar, AI Engineer Summit, YouTube, Wharton, Stanford.

Universal patterns:

• Open with prompt-vs-context contrast
• Teach the write/select/compress/isolate quartet
• Include a hands-on “spec / PRP / plan document” exercise
• Use “vibe coding fails” as the cold-open hook

The gaps (where we play):

1. Mixed-audience. Almost every offering picks a side (consumer or engineer).
2. Mental model, not toolkit. The frameworks above are usable without Python — but no one teaches them that way.
3. Beyond agents. Apply context engineering to research, writing, analysis, document Q&A — workflows everyone has.
4. Debugging context. “Look at this failing prompt, diagnose the context problem.” Nearly absent.
5. A fresh hook. The “vibe coding” opener is exhausted. Open with a real-world non-coding failure case.
6. Stand-alone, not a course funnel. Most free webinars are loss leaders for paid cohorts. Ours can just be the thing.

Open questions for the design phase

The research surfaces design choices we’ll need to make explicitly:

1. How agent-y do we go? Stay in chat-UI workflows (ChatGPT, Claude, Cursor used as chat) or go up to “and here’s what happens when you build an agent?” Affects audience fit.
2. How software-engineering-flavored do we make the framing? The user’s original insight — “SE has discovered techniques that translate” (specs, sub-agents, skills) — is a strong angle. But the audience won’t all share that vocabulary. Translate or assume?
3. One framework or two? Use just Lance Martin’s four strategies (write/select/compress/isolate) — cleaner — or also bring in Breunig’s failure modes? The failure modes are arguably more teachable because they’re concrete and diagnostic.
4. Live demo or pre-recorded examples? Live demo on a real Claude/ChatGPT instance is high-engagement but high-risk. Pre-recorded with live commentary is safer.
5. Take-home artifact? Spec template (à la Cole Medin’s PRP), a context-engineering checklist, or just the slides + reading list?

Sources to drop in the deck

If we end up linking outward, the smallest set that covers the territory:

1. Anthropic’s “Effective Context Engineering” blog — institutional definition
2. Lance Martin’s essay — four strategies framework
3. Drew Breunig’s “Prompts vs. Context” — best non-technical explainer
4. Phil Schmid’s “The New Skill” — best single-link to send attendees afterward
5. Karpathy’s June 2025 tweet — origin moment

---

Next step: Take this synthesis into the design phase — decide the 5 open questions above, then draft the curriculum/outline.