roblox-studio-workflow Checked

e precisely.** List changed Instances/files, modes and client counts, cases

executed, observed Output, failures and fixes, and anything not run. Never translate “looks

plausible” or static inspection into “tested in Studio.”

Execution-location map

| Location | Typical role | Key constraint |

|---|---|---|

| ServerScriptService | authoritative server Scripts/modules | not replicated to clients |

| ServerStorage | server-only assets/data | clients cannot access it |

| ReplicatedStorage | shared modules/assets/remotes | visible to clients; do not store secrets |

| StarterPlayerScripts | client systems across respawns | clone into PlayerScripts |

| StarterCharacterScripts | per-character client behavior | recreated each character |

| StarterGui | authored UI templates | clones into PlayerGui; lifecycle depends on GUI policy |

| Workspace | replicated world and live characters | streaming/ownership may affect client view |

Verify actual Script RunContext and project conventions; class/location shorthand does not excuse

inspection of an existing custom setup.

Reuse before adding

Before creating a RemoteEvent, search by purpose and inspect both endpoints. Before adding a

“manager” or service module, identify the current feature owner. Before introducing folders/tags,

inspect naming and retrieval patterns. Duplicate systems are especially dangerous when both run:

they double-bind input, save twice, create two UI clones, or apply gameplay twice.

Use Attributes for small typed designer-editable metadata that belongs on an Instance. Use

CollectionService tags for discovering sets of Instances across hierarchy. Use folders for

ownership/navigation, not as a substitute for semantic tags. Preserve established schemas and

validate missing/malformed Attributes at system boundaries.

local CollectionService = game:GetService("CollectionService")

local function setupDoor(instance: Instance)
    if not instance:IsA("Model") then return end
    local accessLevel = instance:GetAttribute("AccessLevel")
    if type(accessLevel) ~= "number" then
        warn(`Door {instance:GetFullName()} needs numeric AccessLevel`)
        return
    end
    attachDoorBehavior(instance, accessLevel)
end

for _, door in CollectionService:GetTagged("Door") do setupDoor(door) end
CollectionService:GetInstanceAddedSignal("Door"):Connect(setupDoor)

The owning system must also clean behavior when a tagged Instance is removed/destroyed.

Studio MCP server (optional agent path)

Roblox Studio has a built-in MCP server. When it is connected to your AI client, drive the

same inspect -> edit -> run/verify -> playtest ladder through its tools instead of manual

clicking. This is OPTIONAL: when it is not connected, the manual Studio + Rojo workflow

above and below is the full default path, and nothing in this skill requires MCP, Assistant,

or any other tool. Enable it in Studio under Assistant, then Manage MCP Servers, then

"Enable Studio as MCP server" (it speaks MCP over stdio to the client).

Map the workflow onto the documented tools. Every call also takes a studio_id; use

list_roblox_studios to pick the target when several Studio windows are open.

| Ladder step | MCP tools |

|---|---|

| inspect / inventory | search_game_tree, inspect_instance, script_search, script_grep |

| edit | script_read, multi_edit (creates the script when the path is new) |

| run / verify | execute_luau (needs a datamodel_type: Edit, Client, or Server), get_studio_state, start_stop_play, get_console_output, screen_capture |

| playtest | subagent (types explore, playtest), character_navigation, user_keyboard_input, user_mouse_input |

| Roblox's own docs/skills | http_get (allowed Roblox doc URLs), skill (Roblox-authored reference) |

The discipline does not change with MCP: multi_edit still writes normal, inspectable Scripts

(not a feature hidden in runtime construction), the server stays authoritative, you still edit

source-of-truth files rather than generated output, and you still clean up experiments. Treat

get_console_output and screen_capture as the agent-native evidence for the verification

ladder, and still inspect BOTH server and client output. MCP is a transport for the same

workflow, not a licence to skip it.

Roblox also ships agent-readable docs independent of MCP: an index at

https://create.roblox.com/docs/llms.txt and a .md twin of every docs page (append .md

to its URL). A connected client can reach these through http_get/skill; an offline agent

reads the .md pages directly.

Verification ladder

1. **Static:** project mapping/JSON/XML parses; links resolve; Luau diagnostics/lint if available.

2. **Single session:** correct execution context, hierarchy, no new Output errors/warnings, basic

happy path and cleanup.

3. **Lifecycle:** respawn, reset, reopen/rejoin, streaming or character replacement as relevant.

4. **Server & Clients:** at least two clients for remotes, replication, ownership, player

interaction, joins/leaves, and targeted state.

5. **Device/input:** representative device profiles, orientation, touch, controller navigation.

6. **Stress/failure:** malformed input, spam, empty/full content, rapid transitions, simultaneous

actions, disconnects, and cleanup.

Use the smallest rungs that cover the change, but never substitute a lower rung for a claimed

higher-rung result.

Common failures

|