Pattern #2: From Implementation to Intent

From implementation to intent

Don’t tell people how to do things, tell them what to do and let them surprise you with their results.

– George S. Patton

Talking to machines

We use programming languages to instruct computers what to do. Over the years, that evolved from flipping switches over machine code to yet another new programming language of the day. Compilers help translate that high-level abstraction code construct into low-level machine instructions.

Now, with the advent of LLMs, we can communicate with natural language: using prompts we express what we want. This evolved into a chat interface allowing for a continuous flow of instructions. And the LLM, like a compiler, translates our thoughts into code.

Speclang - End-to-end development in natural language - https://githubnext.com/projects/speclang/

Standard Lib revisited

As developers, we hate to repeat ourselves: similar to shell dot-files, code editors now have a way to configure a system prompt that gets added to all our conversations. Common examples are:  How we want our project directory structure layout, what technology stack it should use, how we want it to write tests and create an openAPI specification for our app.

We can save our own prompts to files and add them as prompt snippets to our conversation. Those snippets are a library we can share for other developers: either local on a project but also across a team or a whole organization or beyond.

Prompt libraries - https://sourcegraph.com/blog/announcing-prompt-library

Specification = Agreement

The great thing about shared and reusable libraries is that they lead to having a conversation. Agile has fostered that conversation on the business level between the customer, product owners and development teams. We focus not only on building things right but also on building the right thing.

The discussion typically creates artifacts such as PRDs (Product Requirements Documents), architectural diagrams, designs and business acceptance criteria. Additionally, security and compliance requirements are formulated and added to it. The process of converging all these requirements is part of our iterative development.

The importance of PRDs in projects - https://www.cursor.new/blog/importance-of-prd-in-cursor-projects

Tests confirm the specification

Tests and examples help a specification by ensuring the system behaves as expected. Ideally they are executable and verifiable to increase the confidence level of the implementation. As the LLM technology is inherently flawed in that it can hallucinate, we really need this as a safeguard.

Describing tests also helps uncover misinterpretations or unclear specifications that might not be evident during the design phase. Language is messy and having clear examples and validation criteria really helps the specification to become clear and focused. 

Applications evolve, and tests help us make sure that when we add new functionality, we can prevent failures and maintain consistency with the original specification without breaking things. Essentially, tests act as a safeguard, ensuring the specification leads to a working product.

Generating tests for test coverage: https://code.visualstudio.com/updates/v1_93#_generate-tests-based-on-test-coverage-experimental

Intent, the root of specification

Human input remains crucial in defining the goals, objectives, and desired outcomes. It still requires a deep understanding of the problem to be solved. Instead of getting bogged down in the granular details of how a system should work, we must focus on articulating the high-level goals and desired behaviors.

This demands a shift in mindset from traditional programming, where the “how” is paramount, to one that prioritizes the “what” and “why.”  In the long run, it fosters collaboration between human creativity and AI’s computational power, leading to better, more innovative solutions. 

In our next AI native development pattern, “From Delivery to Discovery", we’ll describe how AI enables us to learn faster about what we want by discovering different options.