Stage 2: The Prompting Framework

With every asset locked, it's time for Stage 2 — the prompting framework. Before generating a single shot, it's worth pausing here, because this is where most people fail: they overload their prompts with unnecessary information and burn time and credits chasing a result. A typical struggling workflow looks like this — try a prompt, dislike the result, edit it, try again, and repeat until the idea has drifted into a mess that barely resembles the original concept.

To skip that painful loop, the workflow leans on a Claude Skill that packages up the best practices for writing Seedance prompts.

Setting up the skill

Download the skill, then in Claude go to Customize → Skills, hit the plus button, and upload it.

Giving Claude full context

Next, open a fresh chat in Claude and attach two things: the script (as a PDF, telling Claude explicitly that the PDF is the script) and every asset generated in Stage 1 — the main character, the location, both robots, the ball, and the football goal.

To make sure Claude has the full picture, each element is also typed out by name alongside a short description — Adil's civilian top-up outfit, the deserted NYC avenue location at midday, Adil's full exo-armor robot, the opponent robot in black and gunmetal, the regulation football prop, the soda can product prop, and the electric energy-shield goal. The idea is that naming each element this way lets Claude reference them precisely when it writes prompts later.

With the elements listed and the skill installed, the prompting framework is ready, and the workflow moves on to Stage 3 — scene generation.