How to write commentary that describes what actually happened

Last updated 2026-09-04 · 7 min read

What is the short version?

Auto-written sports and gaming commentary has one characteristic failure: it is exciting, well-phrased, and describes a slightly different rally to the one on screen. "Leaping smash, one shot to finish it" reads well and is worthless if nobody leapt and nobody smashed.

1. Go to Vizard Agent and ask for a line on each clip.
2. Read the draft against the footage and say what is wrong.
3. Say what actually happened, and let it verify before rewriting.

What do you need before you start?

The footage and enough knowledge of the sport or game to catch a wrong claim. Vizard Agent can watch a rally and describe it, but you are the one who knows the difference between a drop and a net kill, and that difference is what makes the line worth reading.

• The clips. Cut into their natural units.
• The vocabulary. The words your audience uses.
• Who is who. Names, or shirts, or sides.
• The tone. Professional, or comic commentary.
• Your eye. Someone has to catch the wrong claim.

What do you type into Vizard Agent?

Ask for one description per clip and set the standard in the same breath. The instruction that does the work is not "make it exciting" — it is that every claim in the line has to be visible in that clip.

Prompt

Variants worth knowing:

• "Only what is in the clip." The rule that matters.
• "Check who won the point." The most common error.
• "Name the players." It uses the names you give it.

What does Vizard Agent actually do?

Here is the order Vizard Agent worked in on a real highlights reel after the first set of descriptions came back naming strokes that never happened. What follows is the correction pass rather than the original build, because the correction is where the accuracy actually gets earned.

1. Takes the correction — no leaping smash, no diving save, no kill shot.
2. Re-analyses the rally shot by shot rather than re-phrasing the line.
3. Extracts dense frames at the end of the rally to find the scoring moment.
4. Looks at those frames to see how the point actually finished.
5. Confirms who won it and by what stroke.
6. Rewrites the description to the confirmed facts.
7. Re-cuts the rally where the clip itself was wrong.
8. Checks the frames again for dead time and clean joins.
9. Analyses the next rally stroke by stroke to verify its line too.
10. Views the frames to confirm the stroke type.
11. Regenerates the caption and the matching commentary segment.
12. Compresses the commentary where the corrected line runs long.
13. Rebuilds the audio track so the new line stays in step with the picture.
14. Reviews the finished cut for wording, pace and sync.

Step two is the whole point of this page. When a description is wrong, the tempting fix is to rewrite the sentence — and a rewritten sentence produced without looking at the footage again is simply a second guess. Vizard Agent goes back to the frames.

Step three is how the most common error gets settled. Who won the point is decided in about four frames at the very end of a rally, and the only way to be sure is to sample that end densely and look.

What does the result look like?

A reel where each clip carries a line that a viewer who knows the sport will not argue with. The wording Vizard Agent produces is still styled and still punchy — the correction pass changes the claims being made, not the tone they are made in.

The commentary audio is rebuilt alongside the text, and where a corrected line ran longer than the clip, it was compressed rather than allowed to run over the cut. That is the part that keeps the fix from creating a new problem.

When does this not work well?

Verification needs the evidence to be in the frame. Where the decisive moment happens off camera, or too fast for the footage to resolve, no amount of re-watching settles it and the honest answer is a vaguer line.

• The ball lands out of frame. Nothing to confirm.
• Low frame rates on fast action. The contact is between frames.
• Sports with hidden scoring. Judged events, not decided on screen.
• Very similar-looking players. Say who wears what.
• Claims about intent. "He meant that" is never verifiable.

How do you fix a result that came back wrong?

Name the specific claim rather than saying the line is bad. Vizard Agent keeps the clip boundaries, the frame analysis and the commentary timings, so correcting one claim re-verifies that rally and rewrites one line without touching the others.

• "There was no smash." Re-watched, restated.
• "Wrong player won that." Confirmed from the end frames.
• "That is two rallies, not one." Re-cut, then re-described.
• "The line runs past the clip." Compressed to fit.

How does Vizard Agent compare to doing it yourself?

Writing per-clip commentary by hand is slow but accurate, because you are looking at the clip while you write about it. The failure mode of any automated version is the opposite: fast, fluent, and confidently wrong about which stroke won the point.

Common questions

Why is the first draft wrong so often? Because fluent sports language is easy to produce and specific claims are hard to verify. Vizard Agent needs to be told to check.

Can I just give it the wording? Yes. Vizard Agent uses your lines and times them to the clips.

Does it re-watch every time? When you correct a claim, yes. That is what makes the second draft better.

Can it name the players? Yes. Give Vizard Agent the names and how to tell them apart.

Will it admit it cannot tell? Ask Vizard Agent to. A vague line beats a confident wrong one.

Can it do voice commentary too? Yes. Vizard Agent generates it from the corrected text.

What if a clip contains two rallies? Say so. Vizard Agent re-cuts and then re-describes.

Does correcting one line break the others? No. Vizard Agent rewrites only the one you named.

Can it match a commentator's style? Yes. Name the style and Vizard Agent writes to it.

What about a comic tone? Fine, and Vizard Agent writes it — but the claims still have to be true or the joke does not land.

Does the audio stay in sync? Yes. Vizard Agent rebuilds the track when a line changes length.

Can it work in another language? Yes, including keeping the players' names in their original form.

How many rounds does this take? Fewer than you expect, because each correction is verified rather than guessed.

Why not let it write freely? Because the audience for a highlights reel is people who watched the sport, and they will catch a wrong claim before they finish reading it.