Loading page...
FireFun Games (personal project) · Solo developer & creator
Decebal Cântă — AI-automated history songs for Romanian high-school students
A fully AI-automated content pipeline that turns the Romanian baccalaureate history syllabus into original songs across metal, pop, trap, and manele — with fact-checked lyrics, synced lyric videos, auto-uploaded YouTube content, and matching quizzes.
1 / 2
Decebal Cântă — AI-automated history songs
Decebal Cântă is a personal experiment in what a single developer can build when Claude Code, the Claude Agent SDK, and a stack of AI tools do the heavy lifting. The idea is simple: Romanian high schoolers cramming for the baccalaureate deserve something better than dusty textbooks, so historical figures narrate their own stories through original AI-generated songs across four musical genres. The pipeline is end-to-end automated — from drafting and fact-checking lyrics, through rating and ranking generated songs, to rendering synced lyric videos with Remotion and publishing them to YouTube. The only manual steps are the ones that would break external ToS (pasting into Suno, flipping the YouTube visibility toggle). Everything is released CC0, completely free for students.
Key Features
- Fully AI-automated build flow, from lyrics to YouTube upload
- Four musical genres — metal, pop, trap, manele
- Multi-source fact-checking against textbooks and online references
- Automated song rating, ranking, and rejection
- Word-level lyric timing from WhisperX alignment with Groq Whisper fallback
- Auto-generated quizzes tied to each song, marking which facts are covered
- All content released CC0 and free for students
My Role
- Pipeline architecture
- Multi-agent orchestration
- Next.js frontend
- Remotion video rendering
- YouTube automation
Technical Challenges
- Orchestrating multiple Claude agents as fact-check, rating, and verifier gates
- Getting accurate word-level timings from AI-generated music, with a voice-stem fallback
- Respecting Suno and YouTube terms of service while automating everything else
- Keeping historical accuracy strict without losing musical character
- Making the whole pipeline cheap enough for a hobby budget
Build flow — fully automated with AI
- 01
Cheat sheet per historical figure
Claude Code creates a cheat sheet of facts for each historical figure, making sure every important point from the textbook is covered and fact-checked.
- 02
Three song drafts
Claude Code generates three candidate versions of the song. An in-house skill handles everything end-to-end — I just type a slash command.
- 03
Lyric fact-check
A Claude Code agent fact-checks the song lyrics against multiple textbooks and online sources. Anything wrong or potentially misleading gets scrapped.
- 04
Rate, rank, reject
A Claude Code agent scores the three candidates on catchiness, entertainment, flow, and learning density. Weak ones get scrapped and regenerated; the rest get ranked.
- 05
Lyrics + Suno tags saved
Final lyrics and the matching Suno tags are written to per-song JSON files.
- 06
Next.js tooling UI
A Next.js interface reads those JSON files so I can copy tags and lyrics, and see all three versions side by side.
- 07
Generate in Suno (manual)
I paste the three versions into Suno by hand. This could easily be automated with Claude in Chrome, but that would break Suno's ToS, so I don't.
- 08
Listen and note
I listen to all three versions while doing chores, noting favourites, off-sounding moments, and any creative ideas (which I hand back to Claude to generate more variants).
- 09
Download the winner
I download the winning WAV and drop it into the song directory.
- 10
Second slash command
A single slash command triggers the remaining video and publishing steps.
- 11
Word-level alignment
A WhisperX script in align-only mode extracts accurate word timings using the original lyrics as the reference transcript.
- 12
Alignment sanity check
A Claude Agent SDK agent checks whether the timings make sense, falling back to Whisper on Groq if needed. If timings still look wrong, the flow stops and asks for help — including an alternate path that uses downloaded Suno voice stems.
- 13
Lyric grouping + "did you know"
A Claude Agent SDK agent groups words into logical on-screen lyric blocks and inserts "did you know" popups during instrumental gaps.
- 14
Verifier agent
A second Claude Agent SDK verifier cross-checks the previous agent's output to catch any hallucinated timings.
- 15
Render with Remotion
A script renders the lyric video with Remotion, including lyrics and the "did you know" popups.
- 16
Mux audio
ffmpeg muxes the original audio track onto the rendered video.
- 17
Upload to YouTube
The YouTube Data API uploads the video with its description, tags, category, and metadata pre-filled.
- 18
Go public (manual)
I flip the video to public by hand — YouTube really does not want that automated, for good anti-spam reasons. Fair.
- 19
Site switches to YouTube player
The YouTube link is written back into the song JSON, and the Next.js page automatically swaps the Suno-tags view for an embedded YouTube iframe.
- 20
Auto-generated thumbnail
A thumbnail is generated from the song title and uploaded to YouTube via the API.
- 21
Quiz generation
Claude Code generates a quiz from the core facts in both the song and the cheat sheet, marking which ones are covered by the song and which are not.
- 22
Quiz verification
A Claude Code agent verifies the quiz, its facts, and its answers before shipping.
- 23
Commit, push, deploy
Claude Code commits and pushes to GitHub, which auto-deploys the updated site to Vercel.