As I started using AI more in my work, some of my conversations with Claude and ChatGPT grew to thousands, sometimes tens of thousands, of words. There was often useful research, reasoning and decisions buried in them, but finding something again meant scrolling through long chats full of repeated explanations and approaches I'd already discarded.
I wanted a tool that could take a conversation like that and reconstruct what happened: the questions, the decisions, how the thinking changed and what was worth keeping. That became AI Splicer. I started building it in Lovable on 8 March 2026 and had a working MVP by the end of the day, before returning later to develop it further and rebuild the service on Cloudflare.
The story behind the project
The product
AI Splicer processes long AI conversations and transcripts into structured, reusable knowledge. The aim is not just to shorten a document or list its topics, but to recover the sequence of a discussion, the reasoning behind decisions, useful conclusions and outstanding work.
Users upload text for processing rather than trying to paste an entire long conversation into a single AI prompt. The application divides the material into overlapping chunks, analyses the chunks and then synthesises the results into a more coherent account of the original discussion. It keeps the processing work separate from the browser request so larger documents can be handled as jobs, with progress and results available through the application.
I also wanted people to decide what happened to the output, rather than lock it inside another application. Users could supply their own OpenAI and Anthropic API keys, stored encrypted, so processing ran using their chosen AI accounts. A key feature was the configurable outgoing webhook: once a document had been processed, AI Splicer could send the structured output to an endpoint chosen by the user. That meant they weren't limited to viewing results inside AI Splicer or saving them in Notion. They could route the output into Make, another automation platform, their own application or virtually any system that accepted a webhook. Incoming webhooks also allowed other services to submit documents for processing. Alongside this flexibility, I built a direct Notion API integration for saving results to a database and a Word document export option for keeping a readable report.
The Cloudflare rebuild uses Workers and Queues for the application and processing pipeline, D1 for structured records, R2 for larger source content and Vectorize for semantic retrieval. Document history and search let users revisit processed material. The design was intended to give people several ways to retain and reuse what the AI extracted, instead of treating each upload as a disposable summary.
Building it with AI
AI Splicer was one of two products I decided to build during Lovable's free development day on 8 March 2026. I spent the preceding week preparing the requirements in ChatGPT, asking Claude to challenge them and working on the design before opening Lovable. I prepared a PRD, supporting documentation and an ordered set of development tickets. I had learnt from earlier builds that trying to settle the design after creating all the screens could be unnecessarily expensive.
I set up the two Lovable projects on separate monitors and moved between them while each was building. Most of the planned MVP was in place surprisingly early. For AI Splicer, the initial idea of a simple conversation summary had already grown into a pipeline for ingesting long transcripts, splitting them into chunks, analysing each part and producing a combined synthesis.
The difficulty came when I ran out of prepared tickets but wanted to make the most of the remaining free hours. I started making product decisions while development was already under way, rather than thinking them through beforehand. I kept building, particularly on AI Splicer, and even began addressing security before the day ended. Some of that work was useful, but it made the contrast between planned development and reactive development very obvious.
I refined the Lovable and Supabase application over the following weeks, then returned to it later in 2026 for a Cloudflare rebuild using Claude Code. The migration required more than replacing the database: the pipeline had to preserve the handling of long source text, individual chunk analyses, synthesis, authentication and user-provided API keys. Cloudflare Workers, D1, R2, Queues and Vectorize separated the responsibilities for storage, asynchronous processing and retrieval. The supporting migration work also involved testing and checking that the new application reproduced the existing user journeys.
Results and reflections
AI Splicer reached the stage of a working prototype, with a pipeline for processing long AI conversations into structured analysis. I subsequently rebuilt it on Cloudflare, but I haven't opened it to the public and don't intend to develop it further.
The main reason is that the problem has changed. When I conceived AI Splicer, finding information buried in long AI conversations was frustrating. Since then, the AI platforms themselves have improved their search, memory and knowledge-management capabilities. The standalone product I had imagined is no longer useful enough to justify the additional work required to launch and maintain it.
I still like the concept, especially the flexibility: users could bring their own AI API keys, submit material through webhooks and move the output into Notion, their own automations or a Word document. I also learnt a great deal from building it, particularly about processing long documents in stages, managing asynchronous jobs and handling AI-generated output. It also reinforced the value of preparing requirements before development: once I ran out of planned work during Lovable's free day, I began making decisions reactively simply because I could keep building.
For me, deciding not to progress the product is part of the outcome. A working prototype doesn't automatically justify a public launch, particularly when the underlying user problem is already being addressed elsewhere.
Continue exploring
Interested in how I built it?
The full story covers the initial build, the problems I encountered and what I learnt.
Have an idea or a process you want working?
I take ideas from spec to a working product, and the same approach applies to automation and AI projects inside organisations.


