First Decide What the Text Is Responsible For
| Function of the text | Method and check |
|---|---|
| A shop name, date, message, or other text viewers must understand | Possible treatment: State the exact text, or reserve it for post-production typography What must be preserved: Content, position, and the story moment when it must be readable |
| A page or pattern that only needs to look as though it contains writing | Possible treatment: Use a graphic layout with no legible semantic content What must be preserved: The paper, layout, and approximate visual density |
| A frame that needs no text | Possible treatment: Leave the supporting surface blank and inspect references for embedded text What must be preserved: The object’s existing silhouette, material, and composition |
| Content not yet finalized | Possible treatment: Reserve a place and add it after approval What must be preserved: A clean surface and suitable size relationship for typography |
Separate the story fact from the production method. If the story requires viewers to see “Closed Friday,” preserve that fact. This image pass may use a blank notice card and add the words later. Blankness is only an intermediate production state; the final shot must still communicate the information.
How Do You Rewrite a Prompt with Conflicting Text Requirements?
Original request: At a café entrance, the door clearly says “Closed Friday.” No text of any kind, no subtitles, high quality.
“Viewers must read the notice” conflicts with “no text of any kind.” Choose one of the following versions according to the delivery method.
Option 1: Try to generate the text in this pass
Close view of a café entrance. A pale rectangular notice is fixed at the center of the glass door. The notice contains only the words “Closed Friday.” The entire paper is visible, the camera faces it directly, and these words are the information that must be readable in this shot. Preserve the doorframe and handle positions. No other legible sign text appears in the background, and no subtitle or title is added outside the scene.
Option 2: Add the exact text in post-production
Close view of a café entrance. A blank pale rectangular notice is fixed at the center of the glass door. Its clean surface is fully visible, with no decoration or person covering it. Preserve the doorframe, handle, and reflections in the glass. The notice faces the camera so text can be added in post-production. Do not generate letterforms in the image yet, and do not add subtitles or titles outside the scene.
Record a separate post-production task for option 2: Add “Closed Friday” to the notice, match the paper’s perspective, and ensure the text is already present when the character looks at it. Do not mix that post task into the request for blank paper during image generation.
These are writing examples. For option 1, inspect whether the actual output is correct word for word. For option 2, inspect whether the final composite truly adds the text. Choosing a prompt does not remove the need to check either path.
A readable name is not an available signature
If the story requires recognition of a formal signature, distinguish attribution, a person’s written name, and an approved signature shape. Typing the name does not supply an unavailable handwritten signature.
Suppose a project acceptance sheet must later show a recognizable signature, but the author has specified only who signed. Mark the detail incomplete. With an agreed post-production plan, preserve the signing area, perspective, and light for the approved shape. Do not stage successful recognition of a blank area or let invented strokes fulfill the reveal.
Both direct generation and post-production need exact content first. Listing a compositing task is a plan; its viewing responsibility is delivered only when the final shot actually contains the required readable shape.
Why Does Text from a Reference Remain After the Prompt Says “No Text”?
A prompt expresses the intended output, but a reference may still contain an old sign, number, interface, or subtitle. These are visible image content; a negative phrase does not count as completed cleanup.
Inspect the reference itself before production:
- Locate readable words, letterlike shapes, numbers, title bars, and screen interfaces.
- Decide whether each belongs to the story content being preserved. Verify the version of required text and choose how to exclude anything unnecessary.
- For material you own or have permission to modify, choose cropping, local cleanup, masking, or a different reference according to position.
- Check whether the treatment damages character identity, prop structure, or composition. If it does, change the material or the post-production plan.
If an old café image supplies the doorframe and lighting, its former business hours should not automatically enter the new shot. Clean the old notice first, and limit the reference instruction to doorframe position and lighting rather than copying the whole image.
If you have only a filename or someone else’s written description, request a viewable asset before claiming that the frame contains no text.
When AI Video Adds Subtitles, First Identify the Layer
“Turn off subtitles” may refer to several different things. Inspect the actual file and production steps first:
| Where the text lives | Method and check |
|---|---|
| A separate subtitle or title layer in the editing project | First check: Whether the project contains a selectable text object Treatment: Disable, delete, or edit that layer according to delivery needs, then export again |
| The input image itself | First check: Whether the first frame, character board, or scene reference already includes text Treatment: Treat the reference while preserving required identity and spatial facts, then use it for generation |
| Inside the generated video pixels | First check: Whether the text is already fused into the image Treatment: Evaluate local repair, cropping, or regeneration; before cropping, make sure no character or story information is lost |
| A sign, paper, or screen inside the story | First check: Whether viewers must read it in this shot Treatment: Generate it accurately or add it in post; do not delete it indiscriminately as an unwanted subtitle |
A separate subtitle layer and words embedded in the image require different operations. Exact buttons depend on the tool. Identify the text’s source before finding the corresponding edit control so one subtitle layer does not force a complete redo of a character performance.
If Text Is Added in Post, What Still Needs Checking in the Video?
Text placed correctly on a still image must continue to follow its supporting object in video. Check according to the text’s story function:
- Position: Does it remain attached to the paper, sign, or screen, or drift away when the object moves?
- Readability: When the story requires comprehension, is it blocked by a hand, character, reflection, or frame that is too small?
- Order: If a character acts because of a line of text, has the audience already seen that line?
- Version: Does the notice remain consistent across shots, and if the story changes the wording, does the change occur in the correct shot?
If the text is not final, mark every shot that depends on it as incomplete. Do not write “the character understands the closing notice” in front of a blank sign while leaving the typography task empty.
Text Checklist Before Delivery
- Every required readable text element has an accurate version and a supporting location.
- The team has chosen direct generation or post-production addition rather than requesting legible text and no text at the same time.
- Old text, numbers, and subtitles in input references have been reviewed.
- The final image is inspected again after post-production work, not merely the prompt.
- Irrelevant text does not compete with story information that must be read.
If the issue is confined to a small area of a generated frame, preserve the other correct parts first. Image-to-video prompting remains responsible for action and change in the shot. Text treatment should not casually alter character identity or the story result.
FAQ
Is it always better to add all text in post-production?
It depends on accuracy requirements, image movement, and post-production cost. Story-critical wording benefits from explicit control; background texture that need not be read does not require letter-by-letter typography. Classify each element by function before choosing a production path.
Can adding only “no subtitles” solve every garbled-text problem?
No. Check whether the input contains text, whether the prompt requests it, whether the editing project adds a subtitle layer, and where the problem appears in the actual output. Then apply the correction for that source.
Do it with the skill
In Claude Code, say: “Use /short-drama-image-prompts to inspect the text requirements in these images. Separate text that must be readable, graphic-only writing, blank reserved space, and text to add in post. State exact content and position, and find conflicts between reference images and negative requirements.” In Codex, use $short-drama-image-prompts. Provide viewable references so existing text can be inspected.
Read the method: Method used in this article · Exact text, presentation, and input-reference checks
About Drama SkillsThe short-drama-image-prompts skill on GitHub