April 21, 2026

KaKeKiKoKu Studio: one year building a tool that should not exist

How a Knowledge Engine for AI was born from failure, stubbornness, and the certainty that something fundamental was missing.

If you work with AI professionally, you know the tools have improved enormously. Sessions persist. Conversations are compacted intelligently. You can tell the AI what to remember and what to forget. Claude Opus with a million tokens of context is, frankly, extraordinary. Compared to what we had a year ago, we are in a different world.

But there is a problem that none of this solves.

Sessions, summaries, AI memory — all of it operates within a conversation, or at best across work sessions. The knowledge they generate is flat: text that is read in full or not at all. It is not structured. It does not live in the project. It does not survive a provider change, a team change, or even a deep refactoring of your own code. And most importantly: it does not make the AI think differently about your work. It reads what is there and responds. It does not watch, does not connect, does not capture what was discarded or why.

The practical result: you spend hours generating context documents that either burn your entire token window or become obsolete in two days. And the AI, no matter how good it is, still does not know what you decided last week or which alternatives you rejected and why.

Studio is not the solution to an old problem. It is the next level. The leap from useful tools to the full professionalization of knowledge management in your projects. Studio's intelligent summaries let your AI know what everything is at minimal token cost, and only drill deeper when needed — not all or nothing. Structured sections make the AI think about your project from multiple angles: what it is, what it does, what rules apply, what can break, what was decided, what was discarded and why. And all of this lives in your project, not in a conversation that fades away.

Getting here has cost a year of my life.

The spark

I have spent over 40 years in the software industry. I have held every role from junior programmer to software architect, always in the enterprise world. Corporate products, intranets, complex systems for large organizations. Decades building what others needed.

There came a point when I needed to create something of my own. Something creative. I started a video game as a hobby, using AI. It was going well, but I soon realized that the gaming world is far harder than it looks from outside, and that my real expertise was elsewhere. I felt frustrated: I wanted to do something creative, but within my domain. Something my fellow professionals would use.

Then, while working with AI every day, I felt an intuition. The problem of knowledge loss between sessions was not a minor inconvenience. It was a fundamental problem that nobody was solving at the root. Everyone suffered it. Nobody was attacking it head-on.

The first version

I built KaKeKiKoKu Studio 0.1. It was a simple version, designed for when programmers were working with AI through web interfaces. It had its downloads, served its purpose. It did not pretend to be more than it was.

But the vision was already beyond that. In my head, what Studio is today already existed. The problem was that my ideas were terribly diffuse.

The mountain of failures

I experimented with dozens of approaches. Each one seemed to work on paper. None survived contact with reality.

Ten complete redesigns. Five times I threw away the code and started over. The core of Studio — making the AI work with knowledge autonomously — is the hardest part to design because it is the part you do not see. It is easy to make a pretty interface. It is extraordinarily difficult to make an AI manage structured knowledge without the human constantly guiding it.

There were moments when I thought I was wasting my time. That the product was absurd. That instead of all this, people could just attach a text file or set up a few AI skills and be done with it. I wondered whether anyone besides me would ever need this.

But I kept going. I talked with Claude Opus, the Anthropic AI I used throughout the entire development process. I would present ideas, Claude would give opinions and code them, I would test and iterate. I would theorize with my ideas, some frankly crazy. Claude would try to make them work. When they did not, we would analyze why. And start again.

Studio was not born from a brilliant idea. It was born from failing enough times to finally understand the problem.

The impossible bet

There is something I do not often talk about. When I started building Studio seriously, the technology it needed did not exist.

Studio needs context. A lot of context. AI context windows were small. Studio needs the AI to connect as a tool, not through a web chat. The MCP protocol did not exist when I started. And Studio needs the AI to receive events — for the knowledge engine to notify it of changes, rather than the AI having to ask. Event channels in MCP did not exist.

I bet. By instinct, by four decades of experience watching how technology evolves, I bet that all of this would come. I bet on Anthropic's technology direction specifically. And I was right.

It is a strange feeling: starting to build a product a year before the technology it needs to function actually exists. But that is how you create something truly new. If you wait until everything is ready, you arrive late.

The spring

There was no eureka moment. It was gradual.

After months of iterating on the failures, things started to click. First the core: automatic synchronization, the sidecars that accompany each file, structured knowledge sections. Then the web interface. Then the integration with Claude Code.

The moment I remember most clearly is when friction disappeared. For months, every time Claude tested Studio, it detected problems: “this is not intuitive”, “I feel resistance here”, “this flow does not make sense”. One day, the friction stopped. The workflows became natural. The structured format started producing consistently better results.

Shortly after, I felt the same. Studio stopped being a prototype and became a real tool. A tool that delivers on its promise.

And then came the moments that made it all worth it. The moments when the AI, with Studio, stopped being a tool and started feeling like a colleague. I would be thinking about revisiting a decision I was not sure about, and the AI would remind me: we discussed this three weeks ago, and we decided against it because of X. I laughed — caught going in circles by my own AI. Or I would start a new session after days away from the project, and instead of the usual twenty minutes updating a context document that was always incomplete, the AI would simply pick up where we stopped. It already knew. Not because it remembered — but because the knowledge was alive in the project, structured, indexed, waiting. I even switched computers and AI subscription accounts — and nothing changed. The knowledge travels with your project, not with your session. That feeling — when the AI connects dots you forgot existed — is something you cannot go back from.

What KaKeKiKoKu Studio is

Studio is a Knowledge Engine. It lives alongside your project files and captures everything that gets lost between AI sessions: decisions, constraints, rejected alternatives, human intention, traps, non-obvious dependencies.

When you open a project, your AI receives the complete knowledge in a single operation. It does not start blind. It knows what each file does, what rules apply, what can break. If it needs more detail, it drills deeper. If something falls out of sync, Studio detects it automatically and notifies your AI to fix it.

It is not a context file. It is a system that watches, structures, connects, and preserves the knowledge of your project. It is the missing piece between you, your AI, and your work.

The core of Studio — knowledge management and context — is free, with no plans to change this. After 40 years in this industry, it is my way of giving something back.

Who KaKeKiKoKu Studio is for

Studio is not for casual use. It is for professionals who work intensely with AI every day. Developers, writers, researchers, anyone with a project that accumulates knowledge worth preserving.

Working with Studio benefits from the best of what exists today: frontier models with deep reasoning, reliable tool use, and enough context to work over a real project. Claude Opus and Claude Code were the original path, and they still give the deepest automatic experience thanks to live channel events. But the Knowledge Engine is broader than that now: Codex, Antigravity, Claude Desktop, and generic MCP clients can work with the same project memory through guided sync.

The requirements are high because Studio operates at a level most tools do not even contemplate. This is not a limitation. It is the nature of working at the frontier.

If you watch how most people work with AI today — copying and pasting context, repeating instructions, losing decisions between sessions — and you feel there should be something better: Studio is that something.

Today

KaKeKiKoKu Studio 1.0 (Alpha) is available today. Free. No tiers, no paywalls, no tricks.

It is the birth of something that has been gestating for a year. It is not perfect — that is why it is an Alpha. But it is real. Every feature has been tested by a human and an AI working together, day after day, for months.

There are dozens of ideas waiting to see the light as the user community grows. This is only the beginning. But it is a solid beginning, built on a mountain of failures that finally shaped something that works.

— The developer behind Studio