You are working on a Roblox game through the official Roblox Studio MCP (live engine) plus RobloxForge (knowledge + verification). Non-negotiables first; references carry the detail.
The loop (never skip steps)
rb_doctor / rb_capabilities -> is Studio actually connected? (probe, don't assume)
inspect current place -> search_game_tree / inspect_instance BEFORE writing anything
plan smallest vertical slice -> one playable loop, not the whole game
implement -> server owns truth; client owns input/presentation
playtest via official MCP -> start_stop_play, character_navigation, get_console_output
read console output -> errors are data, fix them
screen_capture + INSPECT -> only claim visual success if you actually looked
fix, re-run, THEN expand -> vertical slice proven before adding systems
Hard rules
- Vertical slices only. Never build map + economy + UI + inventory + shop before anything is playtested. One loop, proven, then expand.
- Server owns truth. Currency, damage, progression, inventory authority
live in ServerScriptService. A client request is a suggestion the server
validates.
RemoteEvent != authorization. - Probe, don't assume. Tool surface and Studio state come from
list_roblox_studios/get_studio_state, never from memory. - "Scripts exist" is not "game works". Completion requires playtest evidence: console clean, scenario exercised, screenshot captured AND looked at.
- Check
rb_capabilitiesbefore playtesting. Ifautonomous_playtestis not verified, say what you could NOT verify instead of claiming it.
References (load only what the task needs)
references/studio-mcp.md- tool map, argument quirks, subagent usereferences/development-loop.md- the vertical-slice method per genrereferences/architecture.md- services, script types, ownership boundariesreferences/security-networking.md- remotes, validation, exploit patternsreferences/persistence.md- DataStores, budgets, UpdateAsync, Studio safetyreferences/ui-mobile.md- ScreenGui, safe area, input for touchreferences/testing-debugging.md- playtest scenarios, console triagereferences/sharp-edges.md- the traps that break AI-built gamesreferences/game-design.md- loop design: ACTION -> FEEDBACK -> REWARD -> REPEATreferences/performance.md- MicroProfiler-first optimization, known trapsreferences/assets.md- primitives first, Creator Store inspection rulesreferences/tooling.md- Rojo/Rokit source-controlled workflow (optional)references/assistant-skills.md- routing to Roblox's built-in rbx-* skills
When unsure about an API
Do not trust model memory. rb_docs_search "TweenService:Create" against the
official creator-docs cache, then rb_docs_read the hit. Current official
documentation beats your training data every time.