Mise en Place for Agentic Coding: Deliberate Preparation as Context Engineering Methodology
This paper proposes "Mise en Place" (MEP), a three-phase preparation methodology that counters the inefficiencies of rapid "vibe coding" by externalizing domain knowledge and structuring specifications to enable effective, parallel agentic coding, thereby introducing "context fluency" as a critical developer skill.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Core Problem: "Vibe Coding" vs. Real Cooking
Imagine you are trying to build a house, but instead of drawing blueprints first, you just start handing bricks to a construction crew and shouting, "Build something cool!" This is what the author calls "vibe coding." It's popular right now with AI coding agents: you type a quick idea, the AI spits out code, you fix the mistakes, and you repeat.
The paper argues this is like cooking without prep. If you start frying a steak but haven't chopped the onions, measured the spices, or sharpened your knife, you're going to spend half your time running around the kitchen frantically searching for things. The result? The food takes longer to make, it tastes inconsistent, and you might burn the steak while looking for salt.
In the world of AI coding, this "running around" happens when the AI writes code that doesn't quite fit your vision. You end up spending more time fixing the AI's mistakes than actually building the product.
The Solution: "Mise en Place" (MEP)
The author proposes a solution borrowed from professional kitchens called Mise en Place (French for "everything in its place"). In a fancy restaurant, before the chef ever turns on the stove, they have every ingredient measured, every tool in its spot, and every recipe memorized. When cooking starts, it's smooth and fast.
The paper suggests we need to do the same for AI coding. Before you let the AI write a single line of code, you must spend time preparing the "kitchen." The author breaks this preparation down into three specific steps:
1. Contextual Grounding (The "Recipe & Philosophy" Phase)
- What it is: You write down not just what you want to build, but why and how it should feel. You turn your "gut feelings" and expert knowledge into written documents the AI can read.
- The Analogy: Imagine you are hiring a sous-chef. Instead of just saying "Make a soup," you hand them a binder that says: "We are making a soup for students. It needs to be simple but deep. Don't use spicy ingredients because our audience is young. Here is the history of this dish."
- Why it helps: AI doesn't have your "gut feelings." If you don't write them down, the AI will guess, and it will likely guess wrong.
2. Collaborative Specification (The "Menu Design" Phase)
- What it is: You have a conversation with the AI to turn your ideas into a detailed design. You tell it what to include, but just as importantly, you tell it what not to include.
- The Analogy: This is like sitting down with a chef to design the menu. You say, "We want a 3-course meal, but skip the appetizer to save time, and make sure the dessert is gluten-free." You are setting the boundaries so the chef doesn't waste time making a 5-course meal or a dessert that can't be eaten.
- Why it helps: It stops the AI from over-engineering or building the wrong features.
3. Task Decomposition (The "Assembly Line" Phase)
- What it is: You break the big project into tiny, independent tasks that the AI can do one by one (or all at once).
- The Analogy: Instead of telling a team of cooks "Make the whole dinner," you give them a list of specific jobs: "You chop carrots," "You boil water," "You plate the fish." Because the jobs are clear and separate, you can have four cooks working at the same time without bumping into each other.
- Why it helps: It allows multiple AI agents to work in parallel without getting confused or stepping on each other's toes.
The Real-World Test: The Hackathon
The author tested this method at a competitive coding event (a "hackathon") where teams had only 5 hours to build a working app.
- The Setup: While other teams started coding immediately (the "vibe" approach), the author spent 2 hours doing the preparation (the "Mise en Place" approach).
- The Result: The author's team used four AI agents working at the same time. Because the preparation was so clear, the agents built the whole app in about 20 minutes of actual coding time.
- The Outcome: The code worked almost perfectly. There were very few bugs, and almost no time was wasted fixing big architectural mistakes. The author's team finished a full educational platform while others were still struggling with basic errors.
The New Skill: "Context Fluency"
The paper concludes that being a good developer in the age of AI isn't just about knowing how to write code or how to type a good prompt. It's about a new skill called Context Fluency.
- The Analogy: Think of an AI agent as a very talented but very literal intern. If you give the intern a vague task, they will do a vague job. If you give them a crystal-clear, detailed, and structured set of instructions, they will be a super-employee.
- The Skill: "Context Fluency" is the ability to take your messy human ideas and turn them into a clean, structured environment that the AI can understand and act on immediately. It's the art of setting the stage so the AI can perform a hit show without you having to run around fixing the lights.
Summary
The paper argues that speed comes from preparation, not rushing. By spending time upfront to organize your thoughts, define your rules, and break down your tasks (Mise en Place), you allow AI agents to work faster and smarter, saving you from the headache of fixing their mistakes later.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.