Agents
JustHold is built with Cursor agents that load skills and rules instead of re-explaining every convention in chat. This page describes that design. The product source is private; the skills catalog publishes the playbooks so others can reuse them.
Skills vs rules
| Kind | Where | Role |
|---|---|---|
| Skills | .agents/skills/ | Task-scoped playbooks agents load when relevant (stack, UI, Supabase, tests, commits) |
| Rules | .cursor/rules/ | Always-on constraints (for example commit narrative style) |
Prefer a skill when the work has a clear trigger (“add an API route”, “write a commit”). Prefer a rule when the constraint should apply even if nobody mentions it.
How we work with agents
- Start from product intent, not from dumping the whole repo into the prompt.
- Let the matching skill supply stack defaults, file placement, and don’t-do lists.
- Keep chat focused on the step at hand; skills own the durable conventions.
- For commits, follow the step-narrative rule: many small commits, each one readable in
git log.
What we publish here
- Skills catalog — map of JustHold-owned skills, each linking to the full live
SKILL.md - Agent skills — open any skill page to read and copy the playbook
- UI patterns — shared product-filter styles
- Testing — why coverage is scoped the way it is
Skill pages load directly from .agents/skills/<id>/SKILL.md, so the docs stay in sync with what our agents use.
Private app, public playbooks
JustHold’s application code is not open source. The agent skills are published for reuse and transparency around how we engineer with Cursor.