Prompt vs context vs harness engineering: how each layer emerged since 2022 and what each one controls when you build with AI coding tools today.
What changes in practice between prompt, context and harness engineering?
Prompt vs context vs harness engineering is the distinction between what you ask a model, what the model can see, and the infrastructure built around the model. Each layer appeared as models changed: prompting dominated from 2022, context engineering followed as windows grew, and harness engineering took over the conversation in 2026.
This article is based on a Full Cycle video, in which Gustavo Dev Doido, the instructor behind the Bootcamp do Dev Doido and the Crazystack Typescript program at Crazystack, walks through the three pillars he considers fundamental to using AI well. The core argument survives the translation: a large language model by itself is a brain that produces nothing. Value comes from what surrounds it.
What is prompt engineering and when did it matter most?
Prompt engineering is the practice of structuring what you ask an AI model so it returns a better answer. It dominated from 2022, when ChatGPT launched with a context window of only 4,000 tokens. With so little room, how you phrased and structured a request carried enormous weight.
The speaker's point is that prompting still matters, but less than it did. If you ask the same thing in N different ways you will get N different answers, and better phrasing still produces more assertive responses. The difference is that modern models tolerate a mediocre prompt and still deliver a usable result. Over the past year, many people noticed they could write a rough request and get good output anyway. That tolerance is a property of better models, not proof that prompting stopped working.
What is context engineering?
Context engineering is the practice of selecting exactly what the model can see inside its limited context window. When Claude Code, Anthropic agentic coding tool, or Cursor, the AI-first code editor, reads your repository, it cannot load the entire codebase, every instruction file, and every dependency at once. Someone has to choose.
The secret, as the video puts it, is that selection itself. You have a world of material that does not fit, so you must fit the right piece at the right moment. There is a real trade-off here: pass too little context and the model invents things; pass too much and it also makes mistakes. The optimum sits in the middle, delivering the right context at the right time.
That optimum is exactly why the third layer exists. You cannot reliably delimit context without tooling that helps you assemble, compress, and inspect it.
What is harness engineering?
Harness engineering is the practice of building the infrastructure around a model: policies, tools, methodologies, and resources that let it work correctly. The term spread through late 2025 and became mainstream in 2026. There is no literal Portuguese translation, and the closest analogy is infrastructure. The field's mantra is blunt: if you are not the model, you are the harness.
A large language model, whether GPT, Opus, or any other, is like a brain that does nothing on its own. It can reason, but it delivers no value until something wraps around it. Harness engineering is what changed that.
Which tools count as a harness?
The existing AI coding tools are harnesses. Claude Code is a harness. Gemini CLI is a harness. Cursor is a harness. Google's Antigravity is a harness. Each one provides resources built around the model: context compaction, planning tools, organization, and visibility into how many tokens the context window has consumed. Remove the tool and only the model remains, and the model cannot do the work.
The layer matters because these tools give you a harness you buy, and you can also build your own on top. The video calls this the difference between the embedded harness and your own harness, and good workflows come from mixing both.
How do you build your own harness?
You build your own harness by configuring the layer between the model and your project. The video lists concrete components, and each one is a file or protocol you control:
A comparison of the three layers
The table below condenses the three layers into the questions they answer and the state of each in 2026.
| Layer | Question it answers | Emerged | Main artifact |
|---|---|---|---|
| Prompt engineering | What do I ask? | 2022, with 4k-token windows | The message itself |
| Context engineering | What can the model see? | As context windows grew | Selected, compacted context |
| Harness engineering | What surrounds the model? | Late 2025 into 2026 | Memory files, rules, skills, MCP, workflows |
One caveat on the underlying hardware of this discipline: as of October 2026, you configure MCP, the Model Context Protocol that lets models call external tools and data sources, inside your own harness rather than inside the model. The protocol is what makes a custom harness interoperable across tools.
Why harnesses matter as software complexity grows
Software complexity grows over time regardless of AI. The video invokes Robert C. Martin's Clean Architecture graph of complexity versus time: code is added, many people pass through the project, and complexity climbs. Architecture and engineering discipline are what keep the project on course.
A harness is how you apply that discipline to AI-assisted development. You configure rules and guidelines for how things must be done, and you configure verification for how the system checks that the work is correct. When a feature is planned according to project needs, implemented, and validated by tests the model was taught to run and create, the harness is what recovers from errors and holds the intended path.
Mixing the embedded harness of a tool like Cursor or Claude Code with your own harness, in blue and red in the original diagram, is what produces a development workflow that survives complexity.
FAQ
- Is prompt engineering still relevant in 2026? Yes, but as one layer among three. Better models tolerate rough prompts, so phrasing matters less than it did in 2022, yet the same request in different forms still yields different answers. Precise prompting still improves results.
- What is the difference between context engineering and prompt engineering? Prompt engineering shapes the message you send. Context engineering selects what fits inside the model's context window, including code, instructions, and references. Prompting is one input; context decides the model's entire field of view.
- What does "harness" mean in AI? A harness is the infrastructure around a model: the policies, tools, memory files, skills, MCP servers, and workflows that let it act correctly. Claude Code, Gemini CLI, Cursor, and Antigravity are all harnesses in this sense.
- Why does passing too much context hurt AI output? Both extremes degrade results. Too little context forces the model to guess; too much dilutes attention and invites mistakes. The goal is the right context at the right time, which only a good harness makes practical.
Turn recorded engineering knowledge into written knowledge
This whole article began as a twelve-minute talk about the three layers of AI work, which is a good reminder that much of the best engineering knowledge now lives in videos rather than text. If you have explanations, lessons, or opinions sitting in YouTube videos, you can turn that recorded knowledge into a structured written article with Skala Blog: paste the video URL, transcribe it, and generate an article that keeps the causal sequence intact.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits