Transformed an idea into a working MVP through research, problem space exploration, and design. Developed a custom prompt engineering framework, crafted user journeys that balanced creativity and usability, and delivered a scalable AI-driven solution.
As outlined in Dan Olsen's The Lean Product Playbook, the initial step in creating digital products is defining the problem space, which consists of three essential components:
My co-founder and I were committed to adhering to Lean Product Principles throughout our journey, beginning with these foundational steps.
Identifying the Ideal Customer Profile (ICP) for Copystream was crucial for directing our focus. We aimed to assist UX Copywriters and Product Designers in crafting production-ready copy directly within design tools. After analyzing various niche data, such as Design and UX team demographics, we determined that our ICP consisted of small to mid-sized businesses (SMBs) in the SaaS sector with Product Design teams of 2-10 members, which may or may not include UX Copywriters.
Our next task was to assess the competitive landscape and calculate the Serviceable Obtainable Market (SAM) using both top-down and bottom-up approaches. Additionally, we brainstormed business models and discussed Go-to-Market strategies.
Through our research and expertise, we pinpointed two major challenges product teams face with copywriting.
First, there are issues with the quality of UX copy. Teams struggle with inconsistencies across product areas, varying writing standards, unclear descriptions and messages, and difficulty in applying specific terms, keywords, and glossaries. There's also a lack of industry-standardized wording.
Second, efficiency and collaboration pose issues. Many teams lack sufficient in-house expertise or are understaffed, leading to lengthy copywriting and review processes. They also deal with unclear iteration history and scattered feedback, alongside the absence of a standardized approach for creating UX copy guidelines.
The final step in defining the problem space was emphasizing the customer benefits, summarized as follows:
Copystream empowers product teams to craft exceptional UX copy from the get-go, all with just a few clicks and no tedious back-and-forth.
With a clear understanding of the fundamentals, we transitioned into the solution space, where my responsibilities included prompt engineering, product requirements, designing high-fidelity mockups, and covering basic brand needs and materials.
Our next step was to collaborate with leading Language Learning Models (LLMs) to produce the core component of our value proposition: production-ready UX Copy.
We chose to work with the widely adopted OpenAI and Anthropic APIs, experimenting with their flagship models, GPT-4o and Claude 3.5 Sonnet. We also tested their more cost-effective versions, GPT-4o Mini and Claude 3 Haiku, since prompt costs significantly impact the business model's viability.
Before diving deeper into API settings like Temperature and Top P, we established a solid prompt engineering framework as our workflow foundation.
Research revealed that a one-size-fits-all solution was nonexistent, with significant variation among existing examples. Consequently, I developed a custom framework, named RIRC.
To maximize model capabilities, we experimented extensively with various prompting techniques, such as Few Shot, Zero Shot, Thought Generation, Self-criticism, and Decomposition.
Additionally, we recognized the need for a more agentic approach to product functionality. During our early access program, a team shared 70 pages of UX Copy guidelines. Inputting such extensive data into prompts increased token costs and affected performance. A "normalizer" agent to distill key information, remove redundancy, apply human feedback, and compress data emerged as an ideal solution.
AI offers a blend of creativity and randomness. For instance, both OpenAI and Anthropic demonstrated impressive vision capabilities, understanding not only text but every UI element in a screenshot, providing high-level context. However, achieving consistent responses when enforcing specific copywriting rules or evaluating existing copy quality based on their Chain of Thought (CoT) proved challenging.
As we evaluated LLM output, I analyzed product requirements and designed high-fidelity mockups to ensure a user-friendly journey.
Building user journeys for AI applications introduced a new interaction paradigm, termed Intent-Based Outcome Specification by Jakob Nielsen, where users specify their desired outcome without dictating the production method.
This new paradigm informed the principles I considered when designing Copystream’s user journey, such as:
These principles support two core user needs:
By this stage, we had completed testing the prompt engineering framework and had all necessary mockups ready. Our priority was to focus on the MVP and present it to real users swiftly.
Developing a Figma plugin was deemed sufficient for the MVP, delaying the actual Platform's development to manage costs and complexity effectively.
Another critical decision was determining which features to include in the MVP. Our thorough testing process provided valuable insights, allowing us to defer features impacted by LLM randomness to future releases and concentrate on those providing immediate user value.
Finally, we chose to proceed with GPT-4o, as it delivered the best results, and a significant price drop occurred at the time. However, we built a proactive and flexible structure to facilitate quick model switching if necessary.
The final phase involved scheduling and launching the Early Access Program, offering a one-month trial of Copystream. My responsibilities included:
The journey from idea to MVP was an exploration of uncharted waters. Fortunately, research serves as a powerful tool in navigating unknowns, yet embracing uncertainty and viewing it as a learning opportunity is crucial.
When building a new product from scratch, aiming for small wins is vital to maintaining momentum and amplifying the sense of progress. Many obstacles will undoubtedly arise. Navigating these challenges until a permanent solution is found is essential for forward movement. After all, the AI landscape evolves rapidly: a problem today might not be an issue tomorrow.
In conclusion, AI represents a superpower that requires harnessing. Combining traditional programming's determinism with AI's productivity can yield impressive results. A balanced approach involves working with AI's traits rather than against them. Imperfection and randomness pose no harm if clearly communicated to users to manage expectations. Conversely, technology capable of processing vast amounts of data swiftly and delivering unparalleled creativity is the way forward.
