Six prompts to remove the AI flavor from content generated by DeepSeek's new model and codex

Sharing 6 new prompts to remove the AI flavor from DeepSeek's new model and codex

Recently I've been thinking about how to use DeepSeek to make code animations like Claude Opus 5.5. This kind of animation can be understood as having the model use code to specify how the text and graphics in the frame should move, and then the program generates them frame by frame according to these rules. It shares a lot of capabilities with making web UI—typography, color schemes, and whitespace all come in handy.

1. Delete duplicate copy

DeepSeek likes to write pages like instruction manuals, and even puts a line of uninformative small English text above the big Chinese headline.

Prompt 1️⃣ Review all the text in the current frame. Delete slogans with no actual information, purely decorative small English text above headlines, and explanations that just restate the meaning of the headline. Say the same thing only once, and prioritize expressing relationships through proximity, grouping, and spacing. When making videos, keep the necessary key text, and don't write explanations of design intent into the frame.

2. Don't add a highlight box to every piece of content

DeepSeek likes to wrap every number and piece of text in a box, adding rounded corners, shadows, and colored borders.

Prompt 2️⃣ Check all the boxes, shadows, rounded corners, and colored borders in the frame. Put related information in the same group together, and make relationships clear through position, spacing, alignment, and thin lines. Remove the layer upon layer of nested boxes, colored borders, and heavy shadows added purely for decoration. Keep clear boundaries only for grouping or for indicating clickable, operable areas.

3. Specify the purpose of each color

Before it was a flood of blue-purple, now everyone's using gradients and semi-transparency

Prompt 3️⃣ First assign fixed colors to the background, body text, secondary notes, highlights, and states. Use the same color for the same meaning, and delete useless decorative colors. Gradients, semi-transparency, and grain can stay, but they must serve the same visual style and must not obscure the body text or steal focus.

4. Unify font sizes, align corners

Specifically for pages where text sizes jump around and icons don't line up with text

Prompt 4️⃣ First narrow the main font sizes down to three or four levels, and use font weight, spacing, and color depth to create hierarchy. Check whether icons look aligned with the text next to them. Check at the final display size whether the text can be read easily, and don't rigidly use 12px small text or 0.5px thin lines. Text in animations must also be legible at normal viewing sizes.

5. One screen, one message

Prompt 5️⃣ First clarify the one thing this screen or this shot most wants people to understand, then order the content. Delete distractions unrelated to the current focus, and don't fill everything at once. On interactive web pages, infrequently used settings can be clicked open later; in videos, secondary information can appear later.

6. Animation & motion design

Prompt 6️⃣ Check each dynamic effect one by one: is it guiding the eye, showing change, advancing the narrative, or providing operation feedback? Reduce purposeless floating, random bouncing, and mechanical character-by-character popping. Keep necessary operation feedback on web pages, and keep camera and graphic changes that advance the narrative in code animations. Adjust speed, pauses, and easing according to the content, and don't apply a uniform 250ms to every motion.

Speaking of which, I can preview the big editing Skill I'm working on. I can already use DeepSeek to edit ten minutes of real-person talking-head footage, along with twenty case videos and screen recordings of operations. I'll have it organized this week!