An Agent Kit for App in Toss Miniapps: Rules Before Vibes

Shipping an App in Toss miniapp looks small until you inventory the checklist: product policy, Korean UX copy, Toss design rules, store assets, console metadata. Ask a bare agent and it usually drops one of those.

I packaged the rules so the agent has to read them. Unofficial kit, works with Codex, Claude Code, or Cursor: agent reads contracts → applies skills → validates results. I used it to push a large batch of miniapps in a day (fifty in the run I wrote about).

Layers

• AGENTS.md — root contract every agent sees first, including routing by task type and “smallest correct change.”
• guide/ — task routers for UI, assets, metadata.
• .agents/skills/ — executable skills: naming, SEO keywords, imagegen, Toss design theme.
• .agents/hooks/ — pre-edit gates and post-edit verification; loops until evidence passes.
• .agents/agents/ — adversarial review subagents.
• .agents/references/ — local references so memory alone is not the source of truth.

When everything lives only in chat memory, concerns bleed. Layers stop that.

The loop

Resolve → read rules → smallest change → strictest gate → on failure replan from root cause → re-run evidence before “done.” The agent is not allowed to declare success until the loop clears. That is how policy violations get blocked early.

Start with a clone, read AGENTS.md, then guide/resolver.md for the task route. The kit is the boring part on purpose: taste and policy first, generation second.