AI Agents Built a Godot Point-and-Click Adventure: What the Act I Devlog Shows About Design, Art and Testing

AI agents built Act I of a Godot point-and-click adventure, and the creator's devlog is unusually specific about how. The creator, Nunc, reports that Claude Code, a Claude Fable subagent, Codex, Gemini TTS and Playwright handled the code, lines, art, voices and play-throughs. The maker kept the decisions. Everything below comes from the maker's own account, so the figures cannot be checked independently.

What is Tihotapska pot, and who made it?

Nunc writes on dev.to and hellonunc.com. The game is Tihotapska pot (The Smuggler's Road). Its hero is Martin, a salt smuggler from the Bloke plateau in 1844, who travels with his mare. Act I is one village and about 20 minutes of play, according to the Act I post.

The maker's summary table for Act I, dated 14 October, lists:

• Engine: Godot 4.7.2, GDScript, single-threaded web build
• Code: 5,299 lines of GDScript, 1,820 lines of Python tools, 155 GUT tests
• Content: 5 rooms, 124 dialogues, 8 items, 772 lines per language
• Voices: 573 Slovenian and 573 English recordings
• Game pack: 63 MB for a Slovenian player, plus a 27.5 MB English voice pack; the engine adds 8 MB
• History: 252 commits, 86 dated decisions, 23 days
• Testers: 67 logged sessions in 38 tester folders

The first article covers the first four days: a 1,472-line game bible, a grey-box prototype, and a voiced demo with 108 commits, 92 tests and 111 spoken lines. The hard part, he says, was a 300-pixel hero whose arms kept switching sides.

Which AI agents did which job?

The Act I post includes a role table. Model names and versions here are as the maker states them.

Against his 551 messages, the agent side used 8.6 million Claude output tokens and 87 Codex sessions. In his words, "the agents propose; the decisions file is mine."

That rule goes back to day one. The first article says page one of the game bible bars the agents from redesigning systems, story or architecture on their own. They have to explain the problem, propose a change and get approval. On day one the agents wrote seven design documents and the maker answered 25 logged decisions.

The set-up has two separations. The model that writes code never writes the lines: Fable gets the script as a separate job. And Codex, from a different vendor, reviews Claude's work, so Claude is not grading itself. That is the maker's reasoning, not a proven result, but it is the most reusable idea in the devlog. For the groundwork of making a project readable to an agent, see our guide to setting up Claude Code for game development.

Where did human judgement change the design?

"Dry, too little soul"

On 26 September, after his fourth play-through, the maker logged one line: "dry, too little soul." The Fable subagent's own review of the script was more precise. Of 28 lines the villagers spoke in that play-through, 22 were orders or next steps, and every character talked like Martin. Fable changed 67 lines that day. Three days later, Fable rebuilt Martin's voice from the 1858 story the game is based on, and 32 more lines changed.

The human noticed the problem; the agent named it.

Monkey Island as a reference, measured in numbers

A separate Claude session took notes on a four-hour walkthrough of the Monkey Island Special Edition. The maker says none of its art, lines or music went into the game. Only the numbers did. In Tihotapska pot, only 11 of 94 dialogues had a choice, the maker says, and players heard 3.7 spoken lines per decision.

The changes that followed: the player picks Martin's lines, used choices dim, a person can end the conversation, and the inn was repainted with a higher camera. Decision 11 had said "no hint system." Decision 62 amended it, so clicking Martin now gives his own line about what is still left to do. The post adds that the staged hints came from a second, deeper review of Monkey Island.

How did the agents handle art and animation?

After the arms episode in the first article, the maker set a rule: nothing is judged on a downscaled screenshot. Every art change is checked on full-resolution crops around the figure. Then a separate Codex run, one that did not paint anything, reviews the result in the running game and saves crops as evidence. When the choice is a matter of taste, the maker picks from a side-by-side page.

The post says Codex's image tool repaints a patch and leaves every pixel outside it unchanged, which suits talk frames, idle frames and door fixes.

Codex also grades its own work inside a painting task. It flagged its four-beat gait for the mare as unacceptable after three tries, and version 3 replaced it. The maker chose that version from a side-by-side, noting version 2 "looked worse." Grok, the third opinion, had lost the one contest it entered by the time of the post.

The pattern is layered review, not trust in one model. Our AI animation editor specimen covers related limits.

What went wrong with the AI voices?

The maker calls the voice work a serial in five episodes.

1. Gemini vs Piper. The first demo used Piper TTS with one Slovenian voice. A blind test of ten lines on 25 September put Gemini TTS against it, and Gemini won every line.
2. Names. The English voice said "Cerknica" three different ways in three takes. Respellings, a pronunciation guide and a light accent all failed a blind test, so the English version got English names (Agnes, Matthias, Gertie).
3. Compression. On 9 October the maker cut the download. Paintings went to lossy WebP (103 to 81 MB), and voices were re-encoded from Vorbis quality 4 to quality 1 (81 to 63 MB). He heard no difference on a comparison page. In play days later, the sound was poor. Measured against q4, the error was only 12 dB below the signal. He restored 916 voices from git, voiced 115 newer lines again, and the build gate now refuses any voice below q4.
4. Transcript checks miss repeats. A take that said "Matija, Matija" where the line had one still scored 0.95. A doubled-word rule followed, and 43 Slovenian and 3 English takes were remade.
5. AI as a listener. Asked to judge quality blind, Gemini found only 1 or 2 of 10 q1 takes. It rated four takes with a deliberate English accent 5 out of 5. The maker concludes it checked words, not sound.

That 81-to-63 MB voice saving was undone. The 14 October table still lists a 63 MB Slovenian pack plus the 27.5 MB English pack, and the post does not break down how that figure is made up.

The maker estimates one voice pass of Act I in both languages at about $0.66 on Gemini and about $5 on ElevenLabs.

For voice and audio work, the lesson is plain: a listening page and an AI listener both passed bad audio, and a numeric gate check now blocks it.

How was an agent-built Godot game tested?

The build gate

According to the post, every commit runs a gate of about 12 minutes:

• Godot engine import
• A content validator that refuses content that could trap a player, such as a choice list with no way out
• A voice check
• 155 GUT tests, including whole-act play-throughs
• Web export and the English voice pack

Agent play-throughs

A Claude agent played the act blind through a browser bridge. Its first run took about 115 actions and found a hidden way out. Codex played both routes end to end in Playwright and wrote a ranked list of findings. To build something similar, see our test plan for agent playtesting in Godot.

Strangers found what the agents missed

On 7 October the maker opened the test page to a Slovenian developer Discord. The first night brought 13 sessions from 9 testers and one finish. Every reporting browser said English, so every session started in English; decision 77 made Slovenian the default.

The second wave, 9 to 12 October, brought 20 sessions from about 18 players. Nine were on phones held sideways, where the game's 30-pixel text shrank to about 9 pixels. Six of fourteen who left the house forgot the neighbour's pot. The fixes: bigger text on small screens, a reminder from Martin at his door, double-click exits, and a half-second click lockout after a room change. Two second-wave players finished, at about 17 and 21 minutes.

The maker writes that the logs said what the comments did not.

A note on Godot web export

The Godot web export docs say web export needs WebAssembly and WebGL 2.0 in the player's browser, and note limits on exporting C# projects, which favours GDScript for browser games.

Who is this workflow for, and what are its limits?

On the devlog's evidence, this workflow suits solo developers willing to:

• write down every decision with a date and their own words
• build an automated gate that blocks bad content and bad audio
• have a different model, ideally from another vendor, review agent art
• put the game in front of strangers early

The limits are real. Only one of a planned five acts is finished, and later acts will be translated into English only when the whole game is done, the maker says.

The ideas worth taking:

1. Keep a numbered, dated decisions file; let agents propose, not decide.
2. Review art at full resolution with a model that did not make it.
3. Measure audio; do not trust a listening page or an AI listener.
4. Recruit real players early and log what they do, with notice.

If you want an agent working inside the Godot editor, start with our godot-ai MCP server install guide. For another long agent-built project and where it fell short, see the Claude-built 3D Terraria rebuild. More projects are in Specimens and under the Godot tag.