When the Name Doesn't Fit

I use cell biology as a naming convention for my AI tools. Every component gets a biological identity: enzymes, organelles, receptors. It started as an aesthetic choice and became a design constraint.

Today I tried to rename my tools from process names to object names. In biology, a tool is an enzyme — an object that catalyzes a reaction. A ribosome does translation. A proteasome does degradation. The tool is the object; the process is what happens when it acts.

Some renames were clean. My emergency-stop tool was called “hemostasis” (the process of stopping bleeding). The enzyme that does this is thrombin. One object, one job. The rename worked immediately — because the tool was well-designed. It does one thing.

Then I hit endocytosis. This tool handles RSS feed scanning, fetching, scoring, sorting, and archiving. In biology, endocytosis involves clathrin coating, dynamin pinching, endosome sorting, lysosomal delivery. Multiple machines, multiple steps. There is no single “endocytosis enzyme” because endocytosis is not one operation — it is a pathway.

I could not name the tool after one object because the tool was not one object. It was five operations wearing a trenchcoat.

This is the insight: when a component resists a name, the naming convention is working. It surfaced a design problem that would have stayed hidden under a process name. “Endocytosis” sounds like one thing. Trying to find the enzyme that does “endocytosis” forces you to see it is actually five things.

The rule I landed on: a well-designed tool maps to one biological object. If no single object fits, the tool needs decomposition, not a creative name.

Three categories emerged from the audit:

Clean renames — the tool IS one object. Hemostasis becomes thrombin. Catabolism becomes proteasome. The rename validates the design.

Decomposition candidates — the tool is an entire pathway pretending to be one tool. These do not get renamed. They get split into focused pieces, each named after the object it maps to.

Already correct — tools that were already named as objects from the start. These had the cleanest implementations. Not a coincidence.

The naming convention is a free architecture review. Every name is a hypothesis: “this component behaves like this biological entity.” When the hypothesis fails — when you cannot find the entity — ask one more question: are all the actions cohesive?

A tool that resists an object name AND bundles unrelated actions needs decomposition. But a tool that resists an object name while all its actions serve one domain is fine as a process name. Endocytosis has four actions — status, fetch, stats, top — all RSS. No single enzyme “does endocytosis,” but the tool is still well-designed because it is about one thing.

The naming test surfaces the question. The cohesion test answers it. Not every process name is a design failure. Some processes genuinely are one thing seen from a distance.

Biology got there first. Cells do not have one protein that “does endocytosis.” They have clathrin, dynamin, Rab5, EEA1, each doing one thing in sequence. But they also have names for the whole process, because sometimes the whole is the right unit of analysis.