AI
Prompt Engineering’s Funeral and the Rise of Context Engineering
Rizel Scarlett opened with a funeral for prompt engineering. Early models like GPT-3 required perfect phrasing because context windows were tiny. That changed after December 2025 when models like Opus reached new capability. “We’re not really just prompting anymore. We’re collaborating with agents,” she said. Scarlett has worked with AI-assisted coding since 2021 at GitHub Copilot. She later led open source DevRel at Block for Goose, the first MCP client and its reference implementation. She now works at Entire.
Context Engineering’s Ambiguity Problem
Context engineering replaced prompts with system information: skills, MCP servers, markdown files like agents.md. But giving too much context causes “context rot” — agents become overwhelmed, hallucinate, forget. Scarlett compared it to reading a 40-page paper and only remembering the last five pages. Teams built workarounds: context compaction, route loops, memory beads, sub agents. Yet the core problem remains ambiguity. The agent still does not understand the user’s intent.
Clarifying Intent with MCP Apps
Scarlett introduced MCP apps as the first primitive for intent engineering. Instead of text, agents render a full UI inside the chat. She demoed with Goose and a neighborhood extension for restaurant ordering. The user clicks a restaurant, sees a menu, and adds items to a cart. Direct interaction eliminates misinterpretation. “We’re aligned on what this means,” she said. The agent understands the exact selection rather than guessing from text.
Confirming Intent with MCP Elicitation
When the agent lacks critical information, MCP elicitation pauses execution and asks the user structured questions. Scarlett showed a Claude Code demo for slide creation. The agent asked “What’s this presentation for?” and “How long do you want this?” instead of guessing. This fills gaps without hallucination. The user provides specific inputs — conference talk, 20 to 30 slides — and the agent proceeds with confirmed intent.
Interpreting Intent with MCP Sampling
MCP sampling lets a server call an LLM to explore multiple interpretations before acting. Scarlett demoed Council of Mind, an MCP server built by a Block teammate. It debates nine perspectives — pragmatist, visionary, optimist — on a question like “tabs or spaces.” Each personality votes and the system selects the best answer. Tabs narrowly won. This provides behind-the-scenes reasoning before committing to an action.
Intent-Based Code Review
AI agents generate code faster than humans can review. Open source maintainers face floods of PRs. Some ignore them, some reject agent PRs entirely. Scarlett proposed intent-based review: capture the agent’s transcripts, prompts, and decisions. Instead of asking “does this code look right?” ask “does the intent match the execution?” Her new company Entire builds this into its workflow, using checkpoints to store the full context.
Q&A
How to prevent prompt injection during intent-based code review? Goose used a GitHub action that scanned PRs for risky content; a red team injection test went unnoticed initially. ▶ Watch (20:35)
How to ensure the client always goes through intent review? Entire captures prompts in checkpoints and extracts the important parts using internal logic. ▶ Watch (22:01)
Notable Quotes
We’re gathered here today to remember our dear hero, prompt engineering. Rizel Scarlett · ▶ Watch (0:30)
It’s not the world we really live in today, especially after December 2025 when the models like got really really excellent with Opus. Rizel Scarlett · ▶ Watch (1:03)
we’re not really just prompting anymore. We’re collaborating with agents. Rizel Scarlett · ▶ Watch (1:11)
I want people to stop talking about intent engineering as though it’s a future concept and start like adapting and working with it now and because it is present tense. You can start using it today. Rizel Scarlett · ▶ Watch (16:52)
if prompt engineering was about saying the right thing, Context engineering was about giving enough information and intent engineering is about making sure the system actually understands you. Rizel Scarlett · ▶ Watch (19:08)
Key Takeaways
- Intent engineering replaces prompt engineering by ensuring the agent understands what you mean, not just what you type.
- MCP apps, elicitation, and sampling provide the primitives for clarifying, confirming, and interpreting intent.
- Intent-based code review shifts focus from code correctness to matching intent with execution.
About the Speaker(s)
Rizel Scarlett is driven by a singular mission: ensuring powerful technology feels human, joyful, and real. As the Tech Lead for Open Source DevRel at Block, she drives technical storytelling for goose, an open source AI agent. Previously at GitHub, she helped devs adopt GitHub Copilot.