Back to the archive

Why heykiku never launched

Eight months, 1,022 commits, zero users. Then five days and a soft launch. A post-mortem.

I spent eight months building heykiku. It came to 1,022 commits, 1,545 issues and pull requests, and about 423,000 lines of code.

Zero users.

The replacement took five days.

Those eight months were mostly weekends.

I started heykiku in January as an AI knowledge base for agencies. An agency could put its knowledge in one place and give people answers scoped to their roles. I believed in the idea. It survived everything that happened next.

The machinery

By September, one development command started the app and six workers: OCR, voice memo, approval, staleness, reconciliation, and insights. I had evaluation lanes, English and Brazilian Portuguese, billing, and a marketing site with more than ten generators for feature-specific social cards. The repo held 855 test files and 90 API routes.

The README still said create-next-app.

I had rules for how the product should work before anyone used it. Analytics needed an EU host and no cookies. I wrote erasure flows and runbooks. I pinned the AI to one Mistral model. I supported two locales from early on. I added MFA, row-level security, Turnstile, and IP restrictions.

Each rule had a reason. Together they gave me months of work that looked like getting ready.

The IP restrictions blocked Polar’s billing webhooks.

A rule I had made to protect the product stopped another part of the product from working. I treated it as one more thing to fix.

The loop

I kept finding things to fix.

The plan documents had names like “review-fixes-v4” and “round-3.” I put the has-plan label on 126 issues. I accumulated 239 agent memory files. I ran multi-model review panels on my own code.

A review found a problem. I wrote a plan. I made the change. Then I reviewed the change and found another problem. The next round had its own document, and the work carried on.

Some of those fixes mattered. Permissions and security findings deserve attention in a knowledge base. But the queue had no end, and I had not put the product in front of a user. Fixes outnumbered features.

I could open the issue tracker and see work waiting for me. I could close a pull request and see the count rise. I had a way to measure how much I was doing each day. I did not have a way to learn whether an agency wanted heykiku as I had built it.

Process felt like progress.

I knew how to move a ticket through a review. Shipping meant exposing the whole thing to someone else.

July

I made 251 commits in May. I made 26 in July.

I do not remember July.

All 26 July commits landed between the 9th and the 16th. They covered a row-level security rollout, GDPR articles, and marketing polish. I opened six new issues. There were no commits in any of my repos during the other three weeks.

The record shows a burst of lockdown work, then nothing. I cannot fill the silence with a better story. I made 62 commits in August and 69 in September. The pace from May never came back.

September

I still had a pre-launch checklist. It had seven items. All seven were unchecked. Webhooks, cron, and email still needed operational work.

On September 15, my last heykiku commit polished analytics telemetry.

I had spent eight months building and hardening the Google Drive importer. I first built it on January 29. After that came re-import staleness, permissions, and security findings. Drive was the flagship integration, and I had disabled it on September 12 while Google OAuth verification was pending.

Google wanted restricted-scope approval for drive.readonly. That meant an unlisted YouTube demo of the full flow. My ticket also said to expect a CASA security assessment, since heykiku stored the tokens on its own server. The verification ticket was still open.

I was not jumping through those hoops. It was not worth it.

I had kept adding safeguards to a product nobody could use yet. Now I faced another process to turn on the integration I had spent months building.

The Drive request was the final straw. I did not want to spend another round getting heykiku ready.

I stopped building heykiku.

Five days

On September 17, I had a whim for a smaller version of the idea. I did not write a plan document. I started.

The old repo had 126 has-plan issues. This time I made a STACK.md with a Rejected list and wrote down the non-goals. Reading those lists now feels like reading heykiku’s autopsy.

I rejected Next.js, Mistral in the running app, Redis, Vercel, and OCR. I left out a chat UI, generated prose, and billing before dogfood. I did not need each part to earn its place in a plan. I needed a reason to add it at all.

Three things stayed. I wanted to manage a knowledge base. I wanted profiles with scoped keys. I wanted notes for workers.

The sharpest non-goal was Google Drive and OAuth. I had just stopped work on heykiku with its Drive integration disabled and a verification ticket open. On day one of Takibi, I refused to build the same dependency into it.

Each rejected piece had a job in heykiku. Together they gave me a product I could keep preparing without ever putting it to use. For Takibi, I wanted the first version to work with the few parts I could finish and try.

I made 97 commits over the weekend of September 19 and 20. Five days after the first commit, I soft-launched on September 22. Anyone could sign up and open the app.

The soft launch does not prove Takibi will work. It does not answer what users will want from it. It means I can start finding out. With heykiku, I spent eight months getting ready to ask.

What survived

I still believe in the agency knowledge base I set out to make. That is why I built Takibi. I stopped believing another review round would tell me when heykiku was ready.

I needed a smaller product that could reach people before I built another system around it.

The idea lived, the architecture died.

heykiku.com now redirects to takibibase.com.