Back to the previous chapter IV - Select the right use case for AI
As you begin to create features using available AI APIs without tools and processes, you may initially find the process straightforward. However, you’ll soon encounter obstacles that can slow down the “Iteration <> Evaluation” loop that is so crucial to nail this empirical development process.
You’ll need to do a few things right to effectively navigate the unique challenges of building with AI.
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/281a9300-4e94-4592-89f8-4147ff5b49a4/Silex_Brand_Symbol.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/281a9300-4e94-4592-89f8-4147ff5b49a4/Silex_Brand_Symbol.png" width="40px" /> Prompts can’t only be available to tech people - everyone helping to build the AI feature needs full access.
</aside>
The challenge with LLMs is that their instructions (prompts) are crafted in natural language. Natural language is much less precise than code, and requires domain-specific expertise that developers may not typically possess. Yet at the same time, these prompts are a critical technical component of AI features.
This creates a situation where:
As a result, it is important that prompts are not embedded directly into code or stored in a generic database.
<aside> 🤓 This pattern is not new.
“Content” was once hardcoded into websites before being abstracted into Content Management Systems (CMS), allowing copywriters (domain experts) to manage content independently of developers.
</aside>
Making them accessible to everyone in a dedicated tool will allow for:
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/c1d180e6-c45b-4dac-ae25-f2f406c6ff77/N12x.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/c1d180e6-c45b-4dac-ae25-f2f406c6ff77/N12x.png" width="40px" /> Seamless collaboration
</aside>
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/a751c16a-bf74-4d5f-b298-c95d991cb642/N42x.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/a751c16a-bf74-4d5f-b298-c95d991cb642/N42x.png" width="40px" /> Composition and personalization
</aside>
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/f76d03bd-da0f-4725-bb2f-9a25a2aae236/N22x.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/f76d03bd-da0f-4725-bb2f-9a25a2aae236/N22x.png" width="40px" /> Effective versioning
</aside>
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/c40a933e-19ec-480e-90d1-42190db328ad/N52x.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/c40a933e-19ec-480e-90d1-42190db328ad/N52x.png" width="40px" /> A/B/N test and canary deployment
</aside>
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/72b006bb-a68f-4dab-a1ab-df4c5e787975/N32x.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/72b006bb-a68f-4dab-a1ab-df4c5e787975/N32x.png" width="40px" /> Smooth rollout and rollback processes
</aside>
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/c40a933e-19ec-480e-90d1-42190db328ad/N52x.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/c40a933e-19ec-480e-90d1-42190db328ad/N52x.png" width="40px" /> Easy audit for compliance
</aside>
Making your prompts accessible is not an end in itself but a prerequisite for effective AI feature management.
<aside> <img src="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/9d7d0f40-a0cc-4145-81da-144542ce396d/Silex_Brand_Symbol.png" alt="https://prod-files-secure.s3.us-west-2.amazonaws.com/d8baae53-7dc0-4bad-a65f-26898d6a633d/9d7d0f40-a0cc-4145-81da-144542ce396d/Silex_Brand_Symbol.png" width="40px" /> Domain experts and product managers are best placed to evaluate the quality of outputs, so they must be easily accessible and continuously reviewed.
</aside>
Your inputs and outputs contain a lot of vital information. They reveal the value you deliver to users, poor results that need improvement, perfect responses that can serve as real-life examples for the AI to emulate, surprising use cases, and ideas for future features.
This data should be easily accessible and manipulable, particularly for domain experts who are best positioned to evaluate the quality of the outputs and product managers who should drive your AI strategy.
<aside> 🤓 Prompts are meta-templates.
Your prompts are essentially meta-templates containing {variables} that are populated with input data from various sources like user interactions, internal databases, or AI outputs from previous prompts.
When these inputs are integrated within your prompt, they create the final prompt sent to the AI to generate an output.
</aside>
