Everyone and their grand dad have access to a chatbot these days, not to mention Donald Trump’s team, allegedly using it to set tariffs for the global economy. Funny that even at that level, the same problem with AI rears its all too familiar ugly head – too generic not good enough! The promise is undeniable, the output, if you dont know what you’re doing (hopefully Donald Trump and co does) will likely be simply mediocre at best. It was this frustration that motivated me to start building my own content writer tool, focusing specifically on overcoming that AI genericity by embedding a deeper understanding of the target audience, their expectations and the writer’s unique insights right into the generation process. The core aim is to enable AI to write as if it’s speaking directly to a specific person, making the final content far more engaging and relatable.
In the spirit of transparency and building in public, this post is the first step in sharing that development journey: I’ll walk you through exactly how the first iteration (V1) works today, openly discuss its current limitations and rough edges, and lay out the vision for where I plan to take it next.
But first, check out Antie Ante:
video transcript:
Auntie Ante was flipping rocks to find the right rock for the hole,
The right rock is sliced with Ante’s spiced knife before 5 mice and a goat,
So auntie Ante calls lost symbols from a soaked rock with 4 ducks and the hole,
The right rock ignites gold lights that spun in a bow,
The right rock is not akin to, it is diamond and gold
So preserve the best stones, the ones with promise and hope,
Thought about unhurried, sealed with three 2s for extra security,
Always, shut the door!
Drive safe home
Iteration 1: The Tool As It Is Today
The cornerstone of this tool, even in its first iteration, is built around what I’d call ‘Persona-Driven Generation’. Instead of just feeding prompts directly into an AI to write an article, the tool first focuses intensely on understanding who the article is for. It analyses the inputs provided – the content brief, any specific instructions, and crucially, example content that reflects what resonates with the target reader – to construct a detailed target audience persona. The rationale behind this initial step is fundamental: by defining this persona upfront, the AI is subsequently instructed to write as if it’s speaking directly to that specific individual. My aim here is to make the generated content feel significantly more personal, relatable, and less like detached, generic AI output. When the model effectively ‘knows’ who it’s talking to, the hope is that the resulting language, tone, and focus become far more clear, targeted and engaging.
Now, regarding the actual implementation of this V1 workflow: it currently lives entirely within a Google Sheet. I opted for this environment primarily for rapid prototyping and the inherent transparency it offered during early development. Building it in a spreadsheet allowed me to quickly map out and test the multi-step logic (persona -> outline -> section writing) and to visually track the inputs and outputs of each distinct stage. Each row represents a content generation task, while the numerous columns meticulously track the process from the initial inputs on the left, through the intermediate AI prompts and responses for persona, outline, and individual headings, moving further to the right. While this layout provides a highly granular view that proved invaluable for debugging and understanding the process flow, it presents its own set of usability considerations, which I’ll delve into later.
To get a copy for yourself, click on the link to view the google sheets based AI writer and make a copy
CLICK HERE TO VIEW TOOL (MAKE A COPY)
So, how does a user actually kick off the process in this V1 spreadsheet? It starts by filling in three key fields for each content task:
- the desired title,
- a content instruction brief outlining the goals and specifics, and crucially,
- Example content showing the style or substance desired.
Once these are provided and the task is initiated (e.g., by clicking a ‘Start’ button or triggering a script), the first major AI step begins: Persona Generation. Here, the tool takes the inputs, particularly the example content, and feeds them into a specific, fairly detailed prompt (starting ‘You are an expert of combined data-driven analysis…’). This prompt instructs the AI to analyse the inputs and synthesize a comprehensive description of the target audience persona.
With the target persona defined, the tool moves to the next phase: Outline Creation. For this step, it gathers all the context it has – the original brief, the example content, potentially insights from search intent/gap analysis if I’ve included that data, and critically, the newly generated target audience persona – and uses this combined information to ask the AI to generate a logical content outline or structure. Once the outline is generated and populated in the sheet, the actual writing begins. The tool doesn’t try to write the whole article at once. Instead, it iterates through each heading defined in the outline. For every heading, it constructs a new prompt telling the AI to write that specific section, ensuring it aligns with the overall brief, the target persona, and its place within the outline structure.
A key aspect I wanted to maintain, even in this spreadsheet version, was transparency and control. As you can likely see in the columns on the screenshot, the actual prompts used for persona generation, outline creation, and each individual heading are displayed directly in the sheet cells. This allows users not only to see exactly how the AI is being instructed but also to manually edit these prompts before execution if they want to fine-tune the instructions for a specific section or the overall task. Furthermore, I included the flexibility to switch between different underlying AI models (like trying Gemini 1.5 Pro). This allows for experimentation to see which model performs best for a particular task or style directly within the workflow.
Iteration 1: An Honest Assessment (Current Limitations & Challenges)
Okay, now for the dose of reality. While V1 successfully demonstrates the core concept of persona-driven generation and allows for some powerful outputs, building it rapidly within a spreadsheet environment inevitably means it has its share of ‘rough edges’. In the spirit of truly building in public, I believe it’s crucial to be upfront about the limitations and challenges I’ve encountered with this first iteration – think of this as the unvarnished truth of a V1 prototype.
The most immediate challenge lies squarely in the user experience, which is inherently tied to the spreadsheet interface. While using Google Sheets was invaluable for my own development and debugging due to the visibility of every step, the result is admittedly complex for an end-user. The sheer number of columns tracking each intermediate prompt and AI response creates significant visual clutter. For someone whose main goal is simply to generate a piece of content, navigating this dense grid, often requiring extensive horizontal scrolling, can feel overwhelming and unintuitive, particularly for those who don’t live and breathe spreadsheets daily.
Beyond the interface friction, I’ve hit some significant reliability and performance bottlenecks. A frequent symptom I’ve observed is the tool failing to properly concatenate the generated sections at the end, sometimes leaving #VALUE! errors or incomplete outputs in the final columns. Based on troubleshooting and analysing the logs my strong suspicion falls on the sheer length of the prompts being sent to the AI, particularly during the section-writing phase. These prompts become very large as they bundle the brief, the detailed persona, outline context, and sometimes lengthy example content. It appears they are frequently exceeding API token limits or processing quotas, triggering the ‘Resource has been exhausted’ errors that halt execution.
Iteration 2: The Vision (Refining and Rebuilding within Google Sheets)
Learning from the V1 prototype is invaluable, and it clearly illuminates the path for the next iteration. My focus for Iteration 2 remains on significantly improving usability, reliability, and performance, while keeping the core strength – the persona-driven approach – intact. However, based on further planning, these V2 improvements will initially be implemented within the Google Sheets environment, refining the existing foundation before a larger leap.
A major focus for V2 usability will be streamlining the user interaction, moving away from navigating the complex multi-column layout of V1 for basic use. The plan is to introduce a new, dedicated ‘Input Form’ tab within the Google Sheet. This form will provide a clean interface where a user simply needs to fill in the three key fields (Title, Content Instruction Brief, Example Content). Once submitted, the entire content generation process (persona, outline, section writing) will run in the background across the subsequent sheets, hidden from the user’s immediate view. Upon completion, the log on the main sheet will be updated with the date, time, a completion message, and potentially a reference to the cell containing the final output for that task, making it much simpler to track and retrieve finished content.
To tackle the critical reliability and performance issues identified in V1, several strategies are planned. Firstly, addressing the prompt length and resource exhaustion: alongside exploring better context summarization, I plan to explicitly add character limit instructions within the prompts sent to the AI. This aims to prevent the generated text for any single step from exceeding Google Sheets’ cell capacity, which might be contributing to errors. Secondly, error handling will be made more robust; the background script will be designed to better catch API errors, timeouts, or quota issues and report these clearly in the log, rather than just resulting in spreadsheet errors. The background processing itself should also improve perceived performance and reliability compared to running complex scripts directly in user-facing cells.
Furthermore, Iteration 2 will introduce significant enhancements to automation and flexibility. To reduce the user’s workload, the new input form will feature optional checkboxes: one allowing the AI to attempt generating suitable Example Content based on the Title, and another allowing the AI to generate the Content Instruction Brief as well. This opens the possibility for users to potentially generate a full draft just by providing a Title and ticking these boxes, letting the AI handle more of the heavy lifting. Addressing the structural rigidity of V1, V2 will also feature Dynamic Headings. Instead of the current fixed number of heading sections, the tool will be instructed to dynamically determine the appropriate number and subject of headings based on the provided context and outline, leading to more naturally structured articles.
These V2 enhancements aim to create a much more user-friendly, reliable, and flexible tool within the familiar Google Sheets environment. The larger architectural step – a complete move away from a Google Sheets front-end UI towards a standalone web application – is currently envisioned for Iteration 3, allowing me to first solidify the core process and features within the current framework.
Join Us on the Journey
So, that brings us to the present moment in this development journey. Iteration 1 exists as a functional, if somewhat unwieldy, spreadsheet-based prototype. Its real success lies in demonstrating the potential of the core idea: using AI to first deeply understand the target audience persona, aiming to generate content that feels more specific and relatable than generic AI output often allows. Looking ahead, the vision for Iteration 2 is sharply focused on tackling V1’s limitations head-on. The plan involves significantly boosting usability and reliability within the Google Sheets environment through a streamlined input form, background processing, smarter error handling, and introducing greater flexibility with features like AI-assisted input generation and dynamic heading structures.
Now, this is where the ‘building in public’ aspect truly comes alive – I need your perspective. This entire process is far more valuable with feedback from people who might actually use such a tool, or who simply find the development process interesting. I would be incredibly grateful to hear your thoughts. Does this core concept of persona-driven AI content resonate with your needs or experiences? What are your biggest frustrations with the AI writing tools you’ve used so far? Considering the V2 plans I’ve outlined – particularly the simplified Sheets form, the options for AI to generate briefs or example texts, and the dynamic headings – what aspects seem most promising? Is there anything crucial you feel is being overlooked? Please don’t hesitate to share any insights, critiques, or ideas you might have.
My commitment is to keep iterating on this tool thoughtfully, driven by the goal of creating something genuinely useful that helps users harness the power of AI to create more effective and relatable content. I intend to continue sharing progress, roadblocks, and discoveries openly as the development of V2 (and eventually V3) unfolds. If you’re interested in following the journey more closely, please subscribe to this blog, subscribe to my Youtube channel, connect with me on LinkedIn, feel free to mention potential future testing opportunities. Thank you for reading and joining me on this venture!
