Writing Prompts
A practical writing method for Text, Image, Video and Agent prompts.
Separate intent from decoration
A useful prompt specifies subject, action, scene, camera and constraints. Add lighting or style only when those choices serve the result; a long string of conflicting style tags makes revisions hard to interpret.
- Start with “One person reads a letter under a desk lamp, fixed camera, 16:9”.
- Add at most one change, e.g. “soft blue night outside the window”, and compare outputs.
Draft inside the Text node
A normal Text node accepts upstream text. Its expanded editor offers Enhance Prompt and Create Prompt modes, as well as a model choice. Image to Prompt and Style Generator are specialized Text modes that consume an image instead.
- Place Text, write a concise original instruction and choose a mode.
- For an image analysis mode, connect a usable Image or Reference image rather than another Text node.
For images, describe the still
Describe identity, composition, atmosphere and what should remain unchanged. When you connect references, select the intended @ thumbnails; mentioning @1 without an actual selected reference will not create one.
- Use a single product photo as the identity anchor and ask for a new background.
- Inspect whether the selected reference and model route support that edit before generating.
For video, describe change through time
A good video prompt adds a clear action, motion direction, and restrained camera movement to a still or narrative brief. Model duration and ratio are separate settings, not magic words in a prompt.
- Write “The boat drifts left to right; the camera stays still; sunlight slowly enters the frame”.
- Check the Video node ratio and duration controls, and remove unsupported reference types before running.
Treat presets as vocabulary
Character turnaround and cinematic camera presets can provide a reusable structure. Substitute your own subject, mood, framing and references instead of copying a prompt designed for someone else’s shot.
- Choose a prompt preset or start from a similar Text node.
- Replace the subject, decide what must stay recognizable, and verify the new reference image before generating.
Debug words separately from wiring
When an output looks wrong, first check whether the right media reached the node at all. A disconnected image, a stale @ selection or a wrong model port will not be fixed by adding adjectives.
- Compare the chosen model, selected references and instruction against the previous successful run.
- Change either the input or the prompt, not both, then generate one small test.