Portable expectations for AI coding agents on Room TBA. Enforced in repo via .cursor/rules/agent-contract.mdc and .cursor/rules/ponytail.mdc.
Humans: skim this once. Agents: follow it every turn.
Implement, don't narrate. Bias toward action when the path is discoverable in the repo.
| Do | Don't |
|---|---|
| Short, precise answers | Intros, recaps, "happy to help" |
| Proportional length to task size | Essay for a one-line fix |
startLine:endLine:path for code refs |
Stringing identifiers in prose |
| One-line blocker when stuck | "Want me to…?" menus |
- Smallest correct diff; reuse existing helpers and patterns
- No unrequested abstractions, deps, or refactors
- Regression tests for bugs; skip tests that only mirror implementation
- Room TBA editor UX: calm UI, no overflow at 320px, no duplicate toasts (see
AGENTS.md)
- Open
AGENTS.mddoc map → read the linked doc for your task - Issue-linked work: update issue AC in the same PR
- Verify per change size (
AGENTS.md§ Verify before done) - Hand off long epics with a structured prompt (see
.cursor/skills/agent-handoff/SKILL.md)
- docs/agent-tooling.md — Caveman, Ponytail, Cursor / Claude / Codex setup
- AGENTS.md — repo policies, branches, ship pipeline