AgentHubv0.4

Skill · Clean code

Loaded by almost every agent. The quality floor for everything AgentHub writes — not a style guide, a set of opinionated choices about what “done” means.

When it activates

Automatically, on every dispatch to backend-specialist, frontend-specialist, debugger, test-engineer, database-architect, and several others. You rarely invoke it yourself.

What it teaches (the short list)

  • No premature abstraction. Three similar lines is fine. Don't extract a helper for two.
  • No defensive code for impossible inputs. Trust internal boundaries; validate at system boundaries only.
  • Comments explain why, not what. Good names replace most comments.
  • No backwards-compat shims unless the data on the other side is out of your control.
  • No half-finished implementations. Complete the feature or revert it.
  • No cleanup you weren't asked for. A bug fix is a bug fix. Not a refactor.

Example: what clean-code refuses

A prompt saying “add a comment explaining this loop” will produce:

// No comment added — the loop's name (chunkByMonth) already describes it.
// Adding "loops through months and chunks entries" restates the code.

The agent will write a comment only if the “why” is non-obvious (hidden invariant, workaround for a known bug, surprising behaviour).

Example: what clean-code produces

Asked to fix a pagination bug:

- const pageSize = 50;
- const results = await db.users.findAll();
- return results.slice(0, pageSize);
+ const results = await db.users.findAll({ limit: 50 });
+ return results;

No added comment, no renamed vars, no extracted helper, no type annotations added “for consistency.” The bug is fixed. Nothing else changes.

Where it lives

skills/clean-code/SKILL.mdis ~1k tokens of rules. Every agent that loads it reads those rules at dispatch time — they're not baked into the model, they're on-demand context.

Pro tips

  • Don't fight it. If an agent refuses to add a comment you asked for, it's not being lazy — it's applying this skill. Rephrase to explain the why.
  • Project overrides go in .hub/instincts.yaml. That's how you encode team-specific rules that should beat this skill.

Next steps