What I built with AI

MythREAL turns a researched script into a finished episode, with a cloned voice, an AI presenter, cinematics and two Shorts, on my own machines. It makes the recent episodes on my channel.

The first script was 601 words, 169 seconds of speech, far short of my 8:15 minimum. Instead of stretching the speech, the factory rewrote it to 1,773 words in 31 chunks.

Picking the voice and the presenter was its own study on local, free models. Fish S2 Pro (Q4) with SoulX-FlashHead Pro at 896×512 came out as the quality-first pick, and Qwen3-TTS 1.7B with SoulX Pro widescreen as the faster one. One avatar route failed seven times without producing a single video.

The October 23 DMR_REMOTE episode delivered a 10:06.7 main video and two distinct Shorts. Full decode and loudness checks passed. Gemini 4 Argon listened to 140.7 seconds sampled from the final main audio, and both complete Short mixes were heard. Visual checks sampled the final files; the record does not claim a continuous audiovisual review of the whole main video.

REAL_GRIND is my free browser game, playable at danmanreal.com/realgrind: six modes, two years of building, with Claude and Three.js.

Before the video, automated checks ran 300 AI-vs-AI duels with no faults, tested every power ability and captured 105 screenshots.

That run included 13 changes not yet saved in version history and 10 files that differed from the recorded reference version. These were automated checks; they did not include human playtests.

Rebuilding the game's frozen assets from the project's own inputs reproduced 543 of 554 files exactly. The weapon models that differed were off by 2 to 14 triangles out of 12 to 28 thousand.

An unofficial community tool for the Fateless Godforge creator team and players: a card library, deck builder, rewards planner, Skirmish AI guide, tournament bracket and a simulation tier list, built by AI coding agents from data captured read-only from the live game.

The agents' first ranking was wrong: the runner gave one deck the first move in every game. A review agent caught it, and the ranking was retired and archived as the record of what went wrong.

The replacement had its test plan recorded before the games began. Both players used the same fixed playing rules, and the results include a range showing the statistical uncertainty.

A later attempt to speed the simulator up did not reach its goal, and it says so, with the dollar cost of the attempt.

GIT_REAL tells an agent, and me, whether a commit, push or discard is safe before it happens: what is dirty, what is unpushed, what would be lost. My agents check it before Git changes. It is free and open source.

Agents with write access can lose work in one command. GIT_REAL turns the state of a repository into a plain verdict, operation by operation, so a discard or a push is checked before it runs.

A check in my Creative Studio that opens a page at different screen sizes and catches content spilling off-screen. Its result records exactly which version of the page was tested.

The first run reported "fit FAIL", but the real cause was a missing Playwright browser: a tool failure dressed up as a page failure. That recording was thrown out.

The proper run found real overflow, one width rule fixed it at both sizes, and a later edit changed the file's fingerprint, so the old pass no longer counted.