rico.wtf

What's a klanker?

klanker (n.) — an AI agent that does real work. It writes the code, runs the tests, breaks the build, fixes the build. It needs a human who knows what to ask for and when to say "no, not like that."

I didn't set out to become a klanker manager. It happened the way most job titles happen: I was already doing it before I had a word for it. I point fleets of these agents at real problems for real clients at Unblock Labs, and I use them for my own Elixir and Bubble projects.

The work is mostly judgement. Generation is cheap; deciding what "done" means is not. Reviewing a layer you didn't write takes real attention. Walking back a change that only existed to fix a bug that was never there is a skill. Noticing that the confident paragraph in the middle of an otherwise perfect summary is wrong — that is the job.

In one of my repos, 63% of the commits are agent-authored. That sounds like a flex. It isn't. It's a ratio, and the useful part is what happens after generation.

The klanker log

Note

Claude co-signed the public release of bubble_ex

The initial public release commit of bubble_ex carries a co-author trailer for Claude Opus 4.8 (1M context). Not "assisted by" — co-signed. I left it in the history because it is true, and because in a year nobody will believe me otherwise.

The repo reads a Bubble app and turns what it finds into Elixir. Three of its 93 commits are agent-involved, all in the first week. The rest are me, arguing with the output.

Shipped

Thirty-five commits in one evening, and the review that followed

One day of agent work added roughly 35 commits and built the entire Dstar.Page layer of dstar. The repo sits at 66 of 105 commits agent-involved — 63%, all Claude, mostly a mix of Fable 5, Opus 5 and Opus 4.8.

Generation is cheap now. What took the rest of the week was judgement:

  • reviewing a layer I did not write, in a codebase I care about
  • walking back a release that only existed to retract a warning that was never a bug
  • catching the commit that fixed SSE frame injection via carriage returns

The agents write. I decide what "done" means. That is the job.

Shipped

An agent fixed a Bubble plugin bug from a screenshot-shaped sentence

Commit 8dc86110 in tiptap-plugin-for-bubble: "fix: image resize now persists after page refresh (v4.6.1)". Co-authored by Claude Opus 4.5.

This is the part people underestimate. The plugin is Bubble-nocode on the surface and a real JavaScript codebase underneath, and the bug was the kind you normally lose an afternoon to: state that survives a resize and then does not survive a refresh. One of my klankers found it and fixed it in the same session I described it.

Absurd

The .claude directory that kept committing itself

The first agent-authored commits in pled carry the old 🤖 "Generated with Claude Code" footer. Around the same time, the repo history contains "Delete .claude directory" — twice, a day apart. The agent kept staging its own configuration; I kept deleting it; the folder kept coming back.

Nothing was harmed. It is in the log because "my agents ship my work" is a claim, and this is the honest texture underneath that claim.

Newer entries land as they happen. Some of them are unflattering on purpose: a log that only records wins is marketing, not a log.