Can a first-time creator get from a blank Baseplate to a game a friend plays, in one sitting?

Roblox has made Studio dramatically more capable for people who already know what to ask it. In April 2026 the Assistant gained Planning Mode, mesh and procedural generation, a playtesting agent and MCP support; Roblox reported that 44 percent of the top 1,000 creators use Assistant or MCP tools, and described the audience as ambitious game developers.¹ The Spring 2026 Creator Roadmap update covers AI, UI styling, asset tooling, analytics, localization and discovery; I could not find an item aimed at someone opening Studio for the first time.²

That leaves a question worth a PM's time: does the agentic Studio work for a first-timer, and if not, where exactly does it break? I wanted an answer grounded in observation rather than opinion, so I used the method a Google or Stripe PM would use on their own product before proposing anything: a friction log.

Three inputs. A stopwatch-timed friction log of my own first hour, with a row at every pause of roughly ten seconds. A coding pass over 33 DevForum threads from first-time and early creators, tagged by funnel stage and theme. A teardown of how Fortnite Creative and UEFN, Meta Horizon, Minecraft, Scratch and the now-closed Rec Room handle the same first hour.

Setup: MacBook, 13-inch screen, Studio installed shortly before on a new Roblox account. Scenario: a first-time creator opens Studio and tries to make and share a small game, using the Assistant when stuck. The Assistant prompt was fixed in advance and written in plain words, so that every later session can be compared against it: “I want to build a rollercoaster ride in a nice amusement park.”

The first build looked like a failure for four minutes, then said it had failed. “Generation is taking longer than expected” sat on screen with an empty viewport, then “The current response exceeded the maximum output length.” Continue rescued it, but a first-timer has no reason to know what output length means or that Continue is safe to press. The result, when it came, was excellent. The problem was the four minutes before it.

Published does not mean shareable, and Studio hands you to a website to find out why. The Publish dialog has no audience field. Share says “Make your game public,” Game Settings says the setting has moved to Creator Hub, and Creator Hub is where the age check and the two-day account rule live. A creator who has just spent half an hour building leaves Studio to learn that nobody can play yet.

Every script action stops for an Accept, and none of them changed anything I could see. Three accept prompts in a row. The checkpoint and kill brick both worked, but from where the camera sat, a working script and a broken one look identical until you press Play. By the third prompt I understood why a first-timer would press Accept All for the session without reading.

What worked, and should be protected: the three-question interview plus the Review card is clear and quick. The finished park was far beyond what I could have built by hand. Play from Studio is instant. The checkpoint script was readable and commented. The Home page does have a first-timer path, the quick tour, but it lives on the Home page and not inside the create flow.

I coded 33 DevForum threads from first-time and early creators, 2023 to August 2026, by funnel stage and theme. DevForum skews toward people who already cleared install and earned posting rights, so true first-hour dropouts are under-represented; Reddit was not accessible from my research setup. With that caveat, the themes line up with the log.

One hypothesis did not survive. I expected blank-canvas paralysis to be the main story. It shows up in only two threads, and in my own run it cost about a minute because the Assistant panel was already open. The expensive breaks come later: while the first generation runs, and after publish.

Problem

A first-time creator can prompt the Assistant, get a genuinely good result, press Play and publish in about half an hour. Along the way the first generation looks broken for four minutes and then reports an error, the build-method question may quietly decide whether the thing they asked for is rideable, script prompts interrupt without visible payoff, and at the end the published link opens for nobody, with the reason on a different website.

Who

Someone opening Studio for the first time who wants to make one small thing and show one friend. Not a returning developer. Not a studio.

Goal

A first-time creator reaches a link one friend can open on a phone in their first session, and learns anything that would stop them from sharing before they spend the hour building.

Proposal A: Publish ends with a shareable link (primary)

• The Publish dialog gains an audience step: only you, friends, everyone. It states plainly who can open the link the moment Create is pressed.
• If the account cannot yet reach friends, the dialog says why (age check, account age) with the date it unlocks, and offers to start the age check from Studio.
• Share never says “make your game public” without a control to do it. It shows the current audience and the next step.
• New creators default to the friends audience once eligible.
• Save to Roblox and Publish to Roblox are visually separated in the File menu, with one line each on what they do.

Proposal B: generation you can watch (secondary)

• Long generations show a stage list in the Assistant panel (ground, decorations, track, cars, script) with the current stage marked, and place each stage in the viewport as it completes rather than at the end.
• When a response hits the output-length limit, the Assistant continues automatically and says so in plain words: “That was a big build. Continuing.” No yellow warning, no unexplained Continue link.
• The first prompt on a new account is framed as a rough first pass with an explicit “this may take a few minutes” before it starts.
• The build-method question shows a thumbnail and one plain sentence per option (“a coaster you can ride, built from parts”; “a sculpted coaster that looks great but does not move”), and if the prompt contains a verb like ride, the Assistant says which options deliver it.

Proposal C: visible confirmation for script changes (secondary)

• When a script lands on a part, the camera frames the part, the part highlights briefly, and a one-line prompt appears: “Press Play and touch it to try.”
• The script-change consent prompt is asked once per session on a new account, with a plain-language explanation, instead of on every action.

Out of scope

Changing the age, ID or account-age requirements. Those are safety policy. This surfaces them earlier; it does not weaken them. Also out of scope: the Data Sharing default in the Publish dialog, which deserves its own review.

Success metrics

Primary

Share of new-account creators with at least one non-creator play within 7 days of first Studio open.

Secondary

Time from first open to first non-creator play. Share of first generations abandoned before completion. Share of Assistant script insertions followed by a Play within 60 seconds. Save vs Publish misclicks in session one.

Guardrails

No change in age-check completion quality. No rise in reports on friends-audience experiences. Assistant cost per new account within budget. No increase in low-effort copies of existing titles.

Risks

Surfacing gates earlier could reduce publishes; the experiment's kill criterion covers this. Auto-continue could hide real failures; it is capped at one continuation before asking. RDC 2026 runs September 10 to 12 and may announce changes here, in which case this becomes an evaluation of what shipped rather than a pitch for it.

Quantitative

Population

Accounts under 7 days old opening Studio for the first time.

Arms

Control. Publish audience step only (A). Generation progress and auto-continue only (B). A and B together. Proposal C rides along with B.

Primary metric

Share with at least one non-creator play by day 7.

Power

Sized once the baseline is known. If the baseline is under 5 percent, expect tens of thousands of first opens per arm, which Studio's volume supports.

Kill criterion

Any arm that lowers the publish rate by more than 10 percent relative without raising the first-player rate is stopped.

Read-out

Day 7 primary. Day 28 for return-to-Studio, to check that a faster first share does not trade against coming back.

Qualitative

Five observed sessions per arm using the same friction-log method, prompt and step list as section 2, so the run on this page is the baseline the arms are compared against. Observer sits beside the participant, does not help, logs a row at every pause of roughly ten seconds, and asks three questions at the end in the participant's own words: the most confusing moment, whether they trusted the Assistant, and whether they would open Studio again tomorrow.

The complete log as I kept it during the run: details block, method, every timed row with its rating and screenshot reference, and the summary. Open the sheet in a new tab.