How to annotate a board game recording with the real moves

Last updated 2026-09-22 · 8 min read

What is the short version?

You have a recording of a game and you want it to feel like a broadcast: a title card, analysis over the top, marks on the good moves. Every one of those depends on Vizard Agent knowing what was actually played, rather than what the video looks like.

1. Give Vizard Agent the recording of the game.
2. Ask it to decode the moves rather than describe the picture.
3. Ask for marks only where the game genuinely turned.

What do you need before you start?

The recording and the framing you want. Vizard Agent reads the board itself, but knowing whose side the video should be from, and whether you want move notation on screen, settles two decisions that are awkward to change later.

• The recording. A GIF, a screen capture, a video.
• Whose side. The board's orientation in the finished cut.
• What to mark. Brilliancies, blunders, the turning point.
• The title. Event, player names, a line for the card.
• Whether you want notation. On screen, or only spoken.

What do you type into Vizard Agent?

Ask Vizard Agent for real moves explicitly. Without that instruction the narration comes out atmospheric and non-specific — "a devastating attack develops" — which is exactly what generic commentary sounds like, and exactly why an audience of players does not believe a word of it.

Prompt

Variants worth knowing:

• "Read the actual moves off the board." The instruction that changes everything.
• "No generic praise." Rules out atmosphere-as-analysis.
• "Where they apply." Marks are earned, not decorative.

What does Vizard Agent actually do?

Here is the Vizard Agent sequence on an 87-second chess recording turned into a properly analysed video. Most of the first half is spent reading the board, and one discovery in the middle of that changes how the rest of the reading has to be done.

1. Opens the recording to see the board, the pacing and the visible moves.
2. Samples several points to reconstruct the flow of the game.
3. Samples the whole sequence, not only the opening, so the analysis covers the finish.
4. Builds a timestamped board contact sheet to trace the exact move order.
5. Decodes the board positions frame by frame so the voiceover can name real moves.
6. Finds the first pass was sampling during the piece animation, and isolates the settled positions.
7. Refines the decoder to separate black and white pieces and recover a clean move list.
8. Extracts the settled positions across the whole recording.
9. Has the footage checked for its turning points so the marks land on real ones.
10. Generates the music bed and the analytical narration.
11. Measures both stems so the narration leads and the music stays under it.
12. Assembles the cut from Black's side with the title and phase cards.
13. Shortens the narration twice so the closing line lands inside the board animation.

Step six is the finding worth carrying to any screen-recorded game. Sampling at regular intervals catches pieces mid-slide, and a board read during an animation is a position that never existed — which produces a move list that is subtly and completely wrong.

Step nine is what keeps the marks honest. A brilliancy mark on an ordinary capture is worse than no marks at all, so the turning points are identified from the game rather than from where the music swells.

Step thirteen is the ordinary discipline of fitting narration to picture. The read is shortened, twice, rather than the board animation being sped up to fit it.

What does the result look like?

A video that names the moves that were played, marks the moments the game actually turned, opens on a title card, and carries a narration you can hear clearly over the music — from the side of the board you asked for.

When does this not work well?

Some recordings cannot be decoded. A board at a steep angle, a heavily compressed capture, or a piece set with unusual graphics leaves the reading unreliable — and unreliable reading is worse than none, because the commentary sounds just as confident.

• Angled or partially hidden boards. No clean grid to read.
• Heavy compression. Pieces blur into their squares.
• Unusual piece sets. The decoder has to learn them.
• Very fast play. Few settled frames between moves.
• Overlays and cursors. They sit on top of the squares.

How do you fix a result that came back wrong?

Quote the move that it got wrong. Vizard Agent re-reads the board around that point, at the settled frames rather than during the animation, and corrects the move list from there instead of rewriting the entire commentary track.

• "That move never happened." The board is decoded again around that moment.
• "It marked an ordinary capture." The turning points are re-identified.
• "The board is the wrong way round." The cut is rebuilt from the other side.
• "The voice is buried." The two stems are measured and the mix rebuilt.

How does Vizard Agent compare to doing it yourself?

By hand you either know the game and write the analysis yourself, which is the honest version, or you write something that sounds like analysis and hope. The second is common and it is exactly what an audience of players is best at detecting.

Common questions

Does it work for other games? Yes. Vizard Agent reads any game whose state is visible — draughts, Go, open-hand card games.

Can it produce the move list on its own? Yes. Ask Vizard Agent for the notation as a file alongside the video.

What if I have the notation already? Give it to Vizard Agent and it checks the recording against it.

Can it explain why a move is good? Yes. Vizard Agent argues from the position rather than offering generic praise.

Will it mark blunders too? Yes, if you ask — Vizard Agent marks what you tell it to look for.

Can the board be flipped? Yes. Vizard Agent can build the cut from either player's side.

What about a live stream recording? Fine, as long as the board is legible and not covered by overlays.

Can it add a title card with players' names? Yes, and it is the cheapest thing to make it look professional.

How long can the game be? Long games work; the reading time scales with the number of moves.

Can it narrate in another language? Yes, with the same decoded move list underneath.

Does the music need to be quiet? Under an analytical read, yes — Vizard Agent measures it against the voice.

Can it produce a short version? Yes. Vizard Agent cuts one to the turning points it identified.

What if the recording is a GIF? That is fine. Vizard Agent reads the frames the same way.

Will it claim a move was brilliant if it was not? Not if the marks come from the game, which is the point of decoding it.