mobilekit

Cross-agent entry point. Claude Code loads SKILL.md; Codex, Antigravity, Cursor, Zed and anything else following the AGENTS.md convention land here.

If you are using mobilekit to build an app

Read SKILL.md now. It holds the phase map, the state model, and the procedure for executing a prompt. Then read prompts/RULES.md — the single source of truth for how prompts are run. Everything below is a map, not a substitute.

The phases are files in workflow/, one per phase — read the one you are running. The library they execute is prompts/.

1-discovery      what are we building        → docs/PRODUCT.md
2-design         screens, navigation, states → docs/DESIGN.md
3-foundation     Expo, NativeWind, domain model, shared components
4-platform       auth, permissions, APIs, database, state, server-side secrets
5-screens        the app itself
6-features       push, media, moderation, payments, AI, i18n, offline, deep linking
7-ship           testing, a11y, performance, store compliance, CI, EAS build
8-observability  consent, analytics, error tracking, what you watch, how you roll back
9-maintain       legacy modernization, SDK upgrades

Slash commands in commands/ are thin aliases for the phase files, for hosts that support them. Codex reads a second set of the same aliases from codex/, because it discovers commands as nested skills rather than as prompt files. Without either, read the phase file directly — that is the whole mechanism, and nothing is lost.

Two things hold regardless of which agent you are:

1. docs/PRODUCT.md gates everything. Absent → run discovery. An app that already ships runs prompts/9-maintain/legacy-modernization.md instead.
2. Everything else you need is in prompts/RULES.md — grill one question at a time, look facts up, request documentation you do not have, invent nothing, keep secrets server-side, report honestly. It is not restated here; a copy would drift.

If you are working on mobilekit itself

This repo is the skill, not an app built with it. See CONTRIBUTING.md.

• prompts/RULES.md is the single source of truth for prompt behaviour; SKILL.md owns the phase model. Everything else points at them.
• Filenames in prompts/ are the ids. A prompt keeps its name wherever it moves; BUILD-PLAN.md entries and the cross-references between prompts depend on it.
• Folder numbers (1-discovery) are phase order and may be reorganized; prompts/README.md, SKILL.md and the workflow/ files carry links to fix — grep -rl 'prompts/' workflow/ SKILL.md AGENTS.md prompts/README.md finds them all.
• A new prompt goes in the folder matching its phase and gets a row in prompts/README.md.
• A new command is written twice: commands/<name>.md for Claude Code, codex/<name>/SKILL.md for Codex. Both hosts read a name: key and want different values in it, so one file cannot serve both — codex/README.md has the detail. install.sh --check fails when one side is missing a command the other has.