Devlog Kit: Your Streams Become Devlogs

Turns a week of stream VODs, voice notes and screenshots on your NAS into a sourced devlog draft, with timestamped highlights, a best-shots album and tagged files.

Comment

Inspiration

I make games and build in public: livestreams, screen recordings, piles of screenshots. Writing the weekly devlog means rewatching hours of footage to remember what I actually did, so it often doesn't get written. All that raw material already sits on storage. I wanted the NAS to do the rewatching for me.

What it does

Devlog Kit takes one folder and a date range and turns that period's stream VODs, voice notes and screenshots into a ready-to-edit devlog:

  • Transcribes the recordings with timestamps.
  • Captions every screenshot and scores it 1-5 for how well it shows progress.
  • Drafts a devlog: summary, what changed, highlights with file name and hh:mm:ss, best screenshots with captions, and "next up" items I said out loud.
  • Saves it next to the footage after I approve the preview.
  • Optionally builds an album of the best shots and tags every cited file devlog.

It only uses what the recordings and images contain. If something was never mentioned, it is not in the draft.

How we built it

It's a Markdown-only orchestration Skill written to the Uliya Skill Development Standard. Each step is a delegate call to a UGREEN AI NAS expert:

  1. file-agent: inventory recordings and screenshots, media durations.
  2. knowledge-agent: audio transcription.
  3. photo-agent: vision model captions, scores and capture time.
  4. document-agent: writing the draft.
  5. knowledge-agent: saving the file.
  6. photo-agent and file-agent: album and tags.

Each step passes only the real result of the previous one. Every write waits for user confirmation and handles waiting_approval. I used Claude Code to read the Skill Developer Package and capability catalog, draft the Skill, and run the kit's static checker.

Challenges we ran into

  • No NAS in Round 1, so every step had to be designed from the capability catalog alone, without seeing real outputs.
  • The kit's template and its static checker disagree on section headings and the capability evidence format. I had to read the checker source to get to a clean pass.
  • Designing honest failure paths: partial transcription failures, no usable screenshots, cancelled approvals, and never claiming a save that did not complete.

Accomplishments that we're proud of

  • A complete six-step workflow that passes the official static checker with zero errors and zero warnings.
  • A no-invention rule built into every step: highlights always point back to a real file and timestamp.
  • Clear limits (8 recordings, 60 screenshots, 12 kept) so runs stay predictable.

What we learned

  • Good orchestration Skills are mostly about contracts between steps: what each expert returns, and what happens when it returns nothing.
  • Keeping every write behind an explicit approval makes a Skill much easier to trust with personal footage.

What's next for Devlog Kit: Your Streams Become Devlogs

  • Validate on a real UGREEN AI NAS: transcription of video containers, vision scoring quality, album and tag writes.
  • Tune prompts on real stream footage.
  • Add output styles for patch notes, Steam announcements and short video scripts.

Built With

  • ai-agents
  • claude-code
  • markdown
  • skills-hub
  • speech-to-text
  • ugos
  • ugreen-ai-nas
  • uliya
  • vision-language-model
Share this project:

Updates

Submission history