Building Roblox Games with Claude
Build a Roblox game in three structured prompts instead of one giant request:
- Prompt 1 (Master Prompt): Full game spec — core loop, zones, progression, enemies/bosses
- Prompt 2 (Bug Fix Pass): Playtest, report every bug/visual issue, let Claude fix them
- Prompt 3 (Depth Pass): Add the "wow factor" — ultimates, cutscenes, summons, unique effects per mechanic
Connect the stack first: Roblox Studio (via MCP) + Claude Code + a 3D asset generator (e.g., ForgeGUI) for models/sounds/VFX that Claude can't natively produce.
Progress:
- Step 1: Pick a viral game to reverse-engineer or reskin (don't clone 1:1 — swap the core mechanic)
- Step 2: Set up the toolchain (Roblox Studio MCP connection, Claude Code, asset-gen tool)
- Step 3: Write the master prompt (use a second AI like ChatGPT to help structure it if needed)
- Step 4: Run Prompt 1, let Claude build the full first version autonomously
- Step 5: Playtest thoroughly — note every bug, broken animation, and missing feedback loop
- Step 6: Run Prompt 2 (bug-fix prompt) with the specific issue list
- Step 7: Playtest again to confirm fixes
- Step 8: Run Prompt 3 to add signature "viral" depth (unique ultimate per ability, boss fights, visual spectacle)
- Step 9: Final playtest across all mechanics/abilities to confirm nothing regressed
- Step 10: Ship / publish and gather feedback
Step 3 Detail: Writing the Master Prompt
The master prompt should specify, in one shot:
- Core loop (e.g., "+1 Mana Every Second" — the addictive incremental hook)
- Hub/world structure (named zones: Spell Hall, Mana Wells, Magic Eggs area, etc.)
- Progression system (trials, unlocks, currency)
- A climactic encounter (named boss, e.g., Elder Treant)
- Theme/reskin twist that differentiates it from the original viral game (spells instead of eyes)
Being this specific up front reduces the number of correction prompts needed later.
Step 5/7 Detail: Playtesting as QA, Not Polish
Playtesting between prompts isn't optional polish — it's the structured feedback input for the next prompt. Treat each playtest as generating a bug list, not a vague impression. Vague feedback ("it feels off") produces vague fixes; specific feedback ("lightning VFX doesn't align with the magic circle spawn point") produces specific fixes.
Example 1: Input: Viral game "+1 Anime Eyes" (simulator: tap to gain eyes, unlock rarities) Output: Reskinned as "Spell Ascent" (+1 Mana Every Second) — same addictive incremental loop, but themed around collecting/casting spells, with a Spell Hall hub, Mana Wells for currency, Magic Eggs for random unlocks, Spell Trials for progression gates, and an Elder Treant boss fight.
Example 2: Input: Post-Prompt-1 playtest reveals spell animations misfire and some mechanics don't trigger Output: Prompt 2 — "Fix the following: [bug list]." Result: stable core loop, ready for depth pass.
Example 3: Input: Game is functional but lacks a "wow" factor for virality Output: Prompt 3 — "Give every spell its own ultimate ability, a cutscene, and a summoned companion." Result: tiered ultimates culminating in a Cosmic Dragon summon, raising the game's shareability.
- Reskin, don't clone. Take the mechanic/loop that makes a game viral, swap the theme — this avoids pure asset-flip and lets Claude generate genuinely different content.
- Use a 3-prompt cadence as a default structure: build → fix → add depth. Resist the urge to cram everything into prompt 1.
- Offload asset generation. Use a dedicated tool (3D models, sounds, VFX) rather than asking Claude to hand-craft assets it's not suited for.
- Let a second AI help draft the master prompt if the spec is complex — clearer prompts reduce rework.
- Always playtest before writing the next prompt. Each prompt should respond to real, observed behavior, not assumptions.
- Save the "spectacle" features for the final prompt — ultimates, cutscenes, and boss summons land better once the core loop is already stable.
- Name everything in the master prompt (zones, bosses, currencies). Concrete names give Claude concrete anchors and produce more cohesive output.
- Don't write one massive prompt trying to get bugs, depth, and core mechanics all correct at once — iterate in passes.
- Don't skip playtesting between prompts; feeding Claude assumptions instead of observed bugs leads to wasted iterations.
- Don't 1:1 clone a viral game's theme — pure clones are less interesting output and more likely to be dismissed as copies.
- Don't rely on Claude alone for 3D models/sound/VFX if a specialized generation tool exists — mixing tools gets better fidelity than text-described geometry.
- Don't add "wow factor" features (ultimates, cutscenes) before the core loop is confirmed stable — spectacle on top of broken mechanics compounds debugging difficulty.