Replace game-like coverage with cinematic direction | Academy

The first full-fight generation has energy, but its camera floats behind Roko and follows the action without making a new visual decision. The result reads like playable coverage: continuous, trailing, and centered on the avatar. The source fixes the grammar of the shot, not the fight choreography.

Exact cinematic-coverage brief

Re-write the fight. @roko alone in the blizzard, samurai coming from all sides, he's cutting them down — parries, counters, black-sand dissolves. Brutal, heavy, real weight in every hit. Remove every 'NO CGI / NOT a game' line and split it into 4 shots: orbit around him mid-fight — low angle, sword slams the ground, shockwave throws them back — close-up, he catches a katana mid-air and counters — wide, one sweep takes out five at once.

This exact string is the source request sent back to Claude or the skill for expansion. It is a directing handoff, not a reconstructed full generation prompt.

Continuous fight generation with a floating camera trailing Roko from behind

One unbroken follow camera stays close to the hero while enemies arrive around him. Repeated bans such as “not a game” keep naming the unwanted reference without specifying a replacement image.

Replace the ban with a picture

The source removes repeated “NO CGI” and “NOT a game” lines. The successful revision names what should be visible instead: a blizzard, attacks from all sides, parries and counters, black-sand dissolves, heavy impacts, and real weight.

This is a directorial test. If a phrase only rejects a category, it does not tell the camera what to photograph. Replace it with light, texture, motion, contact, or composition that could appear in a frame.

The table below is an editorial application of that observed source correction, not a verbatim framework from the presenter.

Rejecting language	Positive directing question
Not a game	Where is the camera, and why does this shot exist?
No cheap effects	What texture, light, debris, and contact should be visible?
Not weightless	What bends, stops, recoils, or travels after the hit?

Give each cut a job

The rewrite breaks the fight into four setups rather than one trailing take.

Setup	Action	Information it adds
50 mm orbit	Roko fights attackers arriving from all sides	Geography and encirclement
24 mm low angle	His sword hits the ground and a shockwave throws enemies back	Mass and scale
85 mm close-up	He catches a katana in the air and counters	Precision and reaction
35 mm wide	One sweep takes out five attackers	Reach and consequence

The lens labels are not decoration. Each framing changes what the edit can prove. The orbit maps the threat, the low angle magnifies impact, the close-up makes a fast counter legible, and the wide shows how far the final action travels.

Before generating, mute the action nouns and read only the camera plan. If every setup would show the same information, the sequence still has one game-like viewpoint split into cosmetic cuts.

The final comparison should answer two questions without relying on taste: does every shot reveal something the previous shot could not, and does the prompt describe the desired frame instead of repeating the unwanted label?