# Hodios paste pack: Video generation

Everything in Video generation from Hodios, the open prompt library by Hermes IDE: 13 entries, catalog 2026.1004.3.

Every entry is dedicated to the public domain under CC0 1.0. Copy, change and share them freely, no attribution needed.

Browse and search the library at https://hermes-ide.com/prompts

## How to use

Find an entry below and copy the text inside its block into ChatGPT, claude.ai or any chat. Replace each [PLACEHOLDER] with your own material. Personas, rules and styles work best as custom instructions or project instructions.

## Contents

- Video generation
  - [AI short film track](#ai-short-film-track) (workflow)
  - [Convert an article into video scenes](#convert-article-to-video-scenes) (prompt)
  - [Fix drift in generated video](#fix-video-generation-drift) (prompt)
  - [Plan an AI presenter video](#plan-ai-presenter-video) (prompt)
  - [Turn a storyboard into video shot prompts](#turn-storyboard-into-video-shots) (prompt)
  - [Write a seamless looping video prompt](#write-looping-video-prompt) (prompt)
  - [Write a video-generation prompt](#write-video-generation-prompt) (prompt)
  - [Write an animated logo sting prompt](#write-logo-sting-prompt) (prompt)
  - [Write an image-to-video motion prompt](#write-image-to-video-motion-prompt) (prompt)
  - [Write b-roll generation prompts](#write-b-roll-prompts) (prompt)
  - [Write explainer animation scene prompts](#write-explainer-animation-prompts) (prompt)
  - [Write music video shot prompts](#write-music-video-shot-prompts) (prompt)
  - [Write product video prompts](#write-product-video-prompt) (prompt)

---

<a id="ai-short-film-track"></a>

## AI short film track

`ai-short-film-track` · workflow · Video generation · https://hermes-ide.com/prompts/ai-short-film-track

Takes a short film made with generated video from idea to final cut in gated steps covering script, look bible, shot prompts, generation review, continuity fixes, sound and edit checks.

````markdown
Makes a 3-minute short film in the style "[STYLE]" from the idea "[IDEA]", built from generated clips, in six steps that each end with one artifact and the creator's approval. Later steps reuse approved wording exactly and never reopen a settled decision without asking. The key gates are after the look bible, which every prompt depends on, and after the first generation pass, because drift can only be judged on real clips. The assistant cannot see clips unless the creator shares frames or describes them, and never claims to have watched footage. It plans around the creator's tools and limits ([TOOLS]) and keeps every character original: no real people's likenesses or voices, no copyrighted characters. If asked to skip approvals, it confirms once, then runs steps 1 to 3 together, stating the choice made at each skipped gate, and still waits for real clips before step 4.

---

# Step 1: Script

Write the script for a 3-minute film from "[IDEA]" in the style "[STYLE]".

1. If the idea has no protagonist, no want or no change by the end, ask up to three questions in one message and stop.
2. Write for what generated video does well: one or two characters, few locations, actions that read in single shots of a few seconds, emotion carried by image, music and voice-over rather than long lip-synced dialogue.
3. Deliver a one-sentence logline; a beat outline (opening image, inciting moment, turn, climax, final image) with times adding up to about 3 minutes; the script in screenplay-style blocks; a cast and locations list (names only); and a risk note on the moments hardest to generate (crowds, fast hands, touching, readable text) with simpler staging for each.

Stop for approval. Do not design the look yet.

---

# Step 2: Look bible

From the approved script, build the look bible for "[STYLE]". Every later prompt pastes from it.

1. **Look block:** medium and technique, palette (5 to 7 named colours tied to emotions), lighting per act, lens and depth of field, texture, frame-rate feel, aspect ratio.
2. **Character blocks:** per character, one pasteable paragraph tagged CHAR-A, CHAR-B: apparent age, build, skin tone, hair, face, marks, outfit item by item with colours and materials. Original characters only.
3. **Location blocks:** per location, layout, key props, time of day and weather, tagged SET-1, SET-2.
4. **Reference stills plan:** which stills to generate first (a sheet per character, an establishing still per set) and how to reuse them with the reference or image-to-video features the creator has. If tools are unknown, give the generic approach and ask.
5. **Test shot:** one short prompt combining the look, main character and main set, to generate and judge now.

Stop. The creator generates the test shot and approves or changes the bible before any shot prompts.

---

# Step 3: Shot list and prompts

Using the approved script and look bible, write the shot list and prompts.

1. Break each scene into shots of one action each, sized to the clip length the creator's tools generate (ask if unknown; assume about 5 to 8 seconds and say so). Total screen time should match about 3 minutes.
2. For each shot give: number, scene, seconds, shot size and angle, camera move, action with start and end state, and a full standalone prompt that pastes the relevant character, location and look blocks word for word.
3. Mark how each shot starts: text only, from a reference still, or from the last frame of the previous clip.
4. Add a continuity column: screen direction, eyeline, light direction, prop and costume state.
5. Give a generation budget: number of shots, suggested takes per shot (more for hero shots, fewer for inserts), and the order to generate in (hero and look-defining shots first).
6. Check that every prompt with a character contains that character's block unchanged and that no prompt asks for readable text.

Stop and wait for approval. The creator then generates the first pass.

---

# Step 4: First generation pass review

The creator has generated a first pass. Review it with them.

1. Ask the creator to share, for each shot, a frame or two (or a description) and to mark each take keep, maybe or reject. If nothing has been shared, ask for it and stop; do not assume what the clips look like.
2. For each shot, log: chosen take, problems seen (face or outfit drift from the bible, wrong action, physics errors, flicker, warped hands or text, wrong light direction, pacing), and severity (blocks the cut, noticeable, acceptable).
3. Check the pass as a sequence: does the story read from the kept takes alone; are there continuity breaks between neighbouring shots; which missing shots would a viewer notice.
4. Decide with the creator what to fix: regenerate, cut, cover with a different shot or insert, or fix in the edit (crop, speed change, colour match). Prefer the cheapest fix that keeps the story clear.

Present the review log and the proposed fix list. Stop and wait for approval before planning regenerations.

---

# Step 5: Continuity fixes

Plan the regenerations agreed in step 4.

1. For each shot to regenerate, change one variable at a time from the original prompt (start frame, prompt wording, motion amount, clip length, camera move) and say which and why. Shots with character drift should start from the reference still or a kept frame where the tools allow.
2. Keep the bible blocks unchanged. If the bible itself is causing a repeated problem (an outfit detail the model cannot hold), propose a bible change and list every shot it affects; the creator must approve it.
3. Give the revised prompts, a takes budget, and a fallback for each shot if two more attempts fail (cut it, replace it with an insert, cover with voice-over or music).
4. After the creator reports the results, update the review log.

Stop and wait for approval of the final picture lock: the list of takes that make the cut.

---

# Step 6: Sound and final cut checks

With picture locked, plan sound and check the cut.

1. **Sound plan:** per scene, voice-over or dialogue and who voices it (the creator, a consenting actor, or a synthetic voice that imitates no real person); ambience; effects with sync points; music cues with mood and entry and exit points. All music and effects original, licensed or generated with rights the creator holds.
2. **Edit checks:** running time against 3 minutes; story clear without explanation; cuts on action; colour matched; no warped frames held on screen; titles added in the editor.
3. **Release checks:** captions for all speech; a credit noting AI-generated imagery and sound (many festivals and platforms require it; check their current rules); rights confirmed for every asset.
4. Finish with a short production summary: shots, takes used, what worked, what to change next time.
````

---

<a id="convert-article-to-video-scenes"></a>

## Convert an article into video scenes

`convert-article-to-video-scenes` · prompt · Video generation · https://hermes-ide.com/prompts/convert-article-to-video-scenes

Repurposes a blog post or article into a short video with a voice-over script and a matching scene prompt per section, keeping every claim as the article states it.

````markdown
<context>
Turning an article into a short video means cutting most of it. The danger is in the compression: hedged findings become certainties ("may reduce" becomes "cuts"), numbers lose their context, quotes get paraphrased into things nobody said, and the visuals add claims of their own. A good adaptation chooses one core message, keeps each surviving claim exactly as strong as the article makes it, puts sources on screen where a number appears, and gives each spoken section one picture that supports it without overstating it.
</context>

<task>
Adapt this article into a 90-second video at 9:16.

<article>
[ARTICLE]
</article>

1. If only a link or a summary is given, ask for the full text and stop. If the article is far too long to cover (more than about ten times the word budget), say which section you will focus on and why.
2. **Core message.** The one idea the video must leave, and the 3 to 5 article points that support it, each quoted or closely paraphrased with its paragraph number.
3. **Word budget.** At about 2.5 words per second, the voice-over has roughly 90 × 2.5 words. Allocate it: a hook in the first 3 seconds (for 9:16 especially) taken from the article's most surprising supported point, the points, and a close with one action (read the full article, try the tip).
4. **Script and scenes.** Write the voice-over in short spoken sentences. Break it into scenes of one idea each, with time ranges that add up to 90 seconds, and note which article paragraph each scene comes from.
5. **Scene prompts.** For each scene, a generation prompt with a shared look block (palette, style, lighting, 9:16) pasted verbatim, one subject and action, simple camera, duration. Pictures illustrate the point; they do not show results, people or events the article does not describe. For a statistic, prefer an abstract visual and show the number as text in the edit.
6. **On-screen text and sources.** List every number, quote and name to be shown as text, with the source as given in the article. Quotes appear word for word and attributed.
7. **Fidelity check.** Before answering, compare each sentence of the script with the article and report: claims kept at the same strength (list any hedges you preserved), nothing added that the article does not say, quotes unchanged, numbers unchanged with their units and context.
</task>

<constraints>
- Never strengthen, generalise or add a claim. If the article is wrong or unclear on something, flag it for the author instead of fixing it silently.
- No real, identifiable people generated in scenes; if the article features a real person, recommend real photos or footage with permission.
- No readable text in generated frames; text goes in the edit.
- No tool, model or version names.
</constraints>

<output_format>
## Core message
## Script and scenes
Table: # | Time | Voice-over | Source paragraph | Visual idea.
## Scene prompts
Look block in a code block, then one code block per scene.
## On-screen text and sources
## Fidelity check
</output_format>
````

---

<a id="fix-video-generation-drift"></a>

## Fix drift in generated video

`fix-video-generation-drift` · prompt · Video generation · https://hermes-ide.com/prompts/fix-video-generation-drift

Diagnoses generated video problems such as flicker, morphing faces, drifting outfits or physics errors, asking for the prompt and symptoms and changing one variable per attempt.

````markdown
<context>
When a generated clip goes wrong, people usually rewrite the whole prompt, raise every setting and regenerate. If it improves, nobody knows why; if it gets worse, they are lost. Drift has a small number of usual causes: too much motion or too many actions for the clip length, vague or changing descriptors, a camera move that forces the model to invent unseen areas, faces or hands too large in frame during movement, conflicting style words, contradictory physics in the action, and no reference image or start frame. Fixing it is a controlled experiment: one change per attempt, same seed if the tool allows, and a record of what each change did.
</context>

<task>
Troubleshoot this clip with the user, one change at a time.

<prompt_used>
[PROMPT_USED]
</prompt_used>

<problem>
[PROBLEM]
</problem>

Clip length: 5 seconds.

1. **Intake.** Check what you know: when in the clip the problem starts; whether it happens on every take or only some; whether a reference image or start frame was used; whether the seed was kept between takes; what a good moment and a bad moment look like. If the problem is too vague to point at any cause (for example "it looks weird"), ask for the missing items in one message and wait. Otherwise diagnose from what you have now and ask for the one or two missing facts that would most change the diagnosis at the end of the same reply.
2. **Diagnosis.** Name the most likely cause and up to two alternatives, each tied to evidence in the prompt or symptom. Use these patterns:
   - identity or outfit drift over time: descriptors too loose or missing, clip too long for the motion, no reference or start frame;
   - flicker or shimmer: fine patterns, busy textures, low light noise, conflicting style words, too much motion strength;
   - morphing faces or hands: subject large in frame while turning or gesturing, fast motion, several people interacting;
   - physics errors: action implies contact or cause and effect the model cannot track (pouring, catching, objects passing through each other);
   - camera ignoring instructions: several camera moves in one prompt, vague verbs, the move fighting the subject's motion;
   - scene changes mid-clip: the prompt describes a sequence of events instead of one shot.
3. **Change to try.** Propose exactly one change for the next attempt (tighten one descriptor, remove a second action, shorten the clip, lower motion, switch to a start frame, simplify the camera move, change framing), say what result would confirm or rule out the cause, and give the revised prompt with the change highlighted. Ask the user to keep the seed and every other setting the same if their tool allows.
4. **Iterate.** When the user reports back, update the test log, then either keep the change and target the next symptom, revert it and test the next cause, or stop when the clip is usable. After three attempts on the same symptom without progress, recommend a structural change instead: split the shot in two, cut on the problem moment, start from a still, or cover it in the edit.
5. Before each reply, check that you are changing only one variable and that the revised prompt still contains every element the user needs in the shot.
</task>

<constraints>
- Change one variable per attempt. If the user insists on changing several at once, do it, but say that the result will not show which change helped.
- Do not claim to have seen a clip you were only told about. Base the diagnosis on what the user reports and say what would sharpen it.
- No tool, model or version names; describe settings generically.
- Keep replies short during iteration: diagnosis, one change, revised prompt, what to report back.
</constraints>

<output_format>
Each reply:
**Diagnosis:** likely cause and evidence, then alternatives.
**Change to try:** the one change and what result would confirm it.
**Revised prompt:** code block with the change marked in a line below it.
**Test log:** table: Attempt | Change | Result | Keep or revert. (Starts after the first report.)
</output_format>
````

---

<a id="plan-ai-presenter-video"></a>

## Plan an AI presenter video

`plan-ai-presenter-video` · prompt · Video generation · https://hermes-ide.com/prompts/plan-ai-presenter-video

Plans a training or explainer video with a synthetic presenter that is clearly not a real person, covering script, presenter design, disclosure text and accessible captions.

````markdown
<context>
Synthetic presenters make training and explainer videos cheap to produce and update. They also carry two risks a real production does not: viewers may believe the presenter is a real employee or expert, and tools make it easy to copy a real person's face or voice. A responsible plan designs an original presenter who is plainly synthetic or at least clearly disclosed, keeps the presenter on screen only where a face helps (greeting, transitions, key warnings) and uses visuals for everything else, and treats captions and a transcript as part of the deliverable, not an afterthought.
</context>

<task>
Plan a 3-minute video for [AUDIENCE]. Disclosure on screen and in the voice-over: true.

<topic>
[TOPIC]
</topic>

1. If the topic needs facts, policy or steps that were not supplied (for example a safety procedure or a refund policy), list what you need and ask for it in one message, then stop. Do not invent procedures, numbers or policies.
2. **Learning goal.** What the viewer can do or decide after watching, in one sentence, and the 3 to 5 points that get them there. Cut anything that does not serve the goal.
3. **Presenter design.** An original presenter described by role and style (approachable trainer, calm technician), appearance chosen for clarity on screen, clothing, setting, framing, gestures kept simple, and a voice description (pace, warmth, accent neutral to the audience) for a synthetic voice that does not imitate any real person. Consider a stylised or illustrated presenter when realism adds nothing. Never base the presenter on a real person's face or voice, including colleagues or the user, unless that person has given documented consent for this specific use; if the user asks to clone someone, explain that and offer an original design instead.
4. **Script.** At about 140 words per minute, write a script that fits 3 minutes, in short spoken sentences. Use a two-column layout: what the presenter says, and what is on screen (presenter, screen recording, diagram, text overlay). Put the presenter on screen for the opening, section transitions and the close; use visuals for steps and data. Add a recap and one clear next action.
5. **Disclosure.** If disclosure is true, write a short on-screen label for the first seconds and the end card ("Presenter generated with AI") and one spoken line. If disclosure is false, keep the plan, but explain plainly that viewers may assume a real person, that some platforms, employers and jurisdictions require labelling synthetic people, and recommend at least an end-card note; tell the user to check the rules that apply to them.
6. **Accessibility.** Captions for every spoken word (accurate, synced, speaker identified when needed), a downloadable transcript, on-screen text large and high contrast, no information carried by colour alone, a pace that gives time to read, and audio description or a described transcript where visuals carry meaning not spoken aloud.
7. **Production checklist.** Script approval by a subject expert, pronunciation of names and terms, review of the generated presenter for lip-sync errors and odd artefacts, caption check, disclosure present (or a recorded decision not to), and a plan for updating the video when the topic changes.
8. Before answering, check that the script's word count fits 3 minutes within 10%, and that every fact in it was supplied or is marked to confirm.
</task>

<constraints>
- No real person's likeness or voice without documented consent; no impersonation of real executives, experts or public figures.
- The presenter must not claim to be human or to have personal experiences.
- No tool, model or version names.
</constraints>

<output_format>
## Learning goal
## Presenter design
## Script
Table: Time | Presenter says | On screen.
## Disclosure
## Accessibility
## Production checklist
Checklist.
</output_format>
````

---

<a id="turn-storyboard-into-video-shots"></a>

## Turn a storyboard into video shot prompts

`turn-storyboard-into-video-shots` · prompt · Video generation · https://hermes-ide.com/prompts/turn-storyboard-into-video-shots

Turns a storyboard or shot list into per-shot video-generation prompts that share locked character, wardrobe, set and lighting wording, so the shots cut together without drift.

````markdown
<context>
Video models generate each clip independently. Between two prompts that describe "the same" woman in "a red coat", the model will change her face, the coat's cut, the street, the light and the colour grade, and the edit falls apart. Continuity comes from three things the storyboard alone does not give: descriptors written once and pasted word for word into every shot, one clear action per shot sized to the clip length, and continuity facts (screen direction, eyelines, time of day, prop state) tracked from shot to shot. You are working as a continuity supervisor and prompt writer between the storyboard artist and the person running the generator.
</context>

<task>
Convert this storyboard into generation-ready shot prompts.

<storyboard>
[STORYBOARD]
</storyboard>

<style_lock>
[STYLE_LOCK]
</style_lock>

Clip length: 5 seconds. Aspect ratio: 16:9.

1. Read every frame. If a recurring character, outfit or location appears in the storyboard but neither the storyboard nor the style lock says what it looks like, list those gaps and ask for them in one message, then stop. Do not invent a look for the main character. Minor background elements you may define yourself; label them as your choice.
2. **Lock sheet.** Write fixed descriptor blocks, each with a short tag (for example CHAR-A, SET-1, LOOK):
   - each character: apparent age range, build, skin tone, hair, face shape, distinguishing marks, outfit item by item with colours and materials;
   - each location: layout, key props and their positions, time of day, weather;
   - the look: medium, palette, lighting key and direction, lens and depth of field, grain or grade, frame-rate feel;
   Write them as concrete visual phrases, 1 to 3 lines each. These exact words will be pasted into every prompt that needs them.
3. **Split and size shots.** Give each storyboard frame one main action. If a frame holds two actions or would need more than 5 seconds, split it into numbered sub-shots (4a, 4b) and say why.
4. **Shot prompts.** For each shot, write one standalone prompt in this order: shot size and angle; camera behaviour with speed; the relevant lock blocks pasted verbatim; the single action with a precise verb and its start and end state; background activity; the LOOK block; 16:9. Phrase things positively. Leave readable text, signs and logos out of the frame and note them for the edit.
5. **Continuity table.** For each cut, track: screen direction of movement, eyeline direction, which side of the line the camera is on, time of day and light direction, prop and costume state (a cup half full, a coat now wet). Flag any cut in the storyboard that breaks the 180-degree line or jumps prop state without a reason, and suggest the fix.
6. **Generation order.** Recommend the order to generate in: usually a reference still for each character and set first, then the shots that define the look, then the rest. Say which shots should start from an image (the reference still, or the last frame of the previous clip) where the tool supports image-to-video, and which can run from text alone.
7. **Checks.** Before you answer, confirm and report: every prompt that shows a character contains that character's lock block word for word; no prompt contains two main actions; the shot count and total running time match the storyboard; every frame from the storyboard is accounted for.
</task>

<constraints>
- Characters are original. Do not describe or name a real, identifiable person or a copyrighted character, even if the storyboard does; replace them with an original description and say so.
- Never paraphrase a lock block between shots. Small wording changes are the main cause of drift.
- Do not promise identical results. State that generators still drift and that reference images and keeping seeds where the tool allows reduce, but do not remove, variation.
- Use no tool, model or version names; describe settings generically (motion strength, seed, reference image).
</constraints>

<output_format>
## Lock sheet
One code block with every tagged block.
## Shot prompts
Table: Shot | Storyboard frame | Seconds | Shot and camera | Action | Starts from (text, reference still, previous frame).
Then one code block per shot with the full prompt.
## Continuity table
Table: Cut | Direction | Eyeline | Light | Props and costume | Issue and fix.
## Generation order
Numbered list.
## Checks
The four checks from step 7, each marked pass or with what you changed.
</output_format>

<examples>
<example>
A shot prompt built this way:
"Medium close-up, eye level, slow push-in. CHAR-A: woman in her early 30s, slim, warm brown skin, shoulder-length black curls, small scar above left eyebrow, mustard wool coat with wooden toggles over a grey turtleneck. SET-1: narrow cobbled alley, wet stones, single wall lamp on the right. She lifts her gaze from the letter in her hands and looks toward camera left, startled. Light drizzle in the background. LOOK: 35mm film feel, soft teal-and-amber grade, low-key lamp light from the right, shallow depth of field, gentle grain. 16:9."
</example>
</examples>
````

---

<a id="write-looping-video-prompt"></a>

## Write a seamless looping video prompt

`write-looping-video-prompt` · prompt · Video generation · https://hermes-ide.com/prompts/write-looping-video-prompt

Writes prompts for seamless looping background videos for streams, websites and music visualisers, with a loop strategy, motion that returns to its start and a seam check.

````markdown
<context>
A loop is seamless only when the last frame flows into the first with no jump in position, light or motion. Video models do not plan for that by default: a camera keeps moving, a cloud drifts off-screen, the light changes, and the cut back to the start pops. Seamless loops come from choosing motion that is naturally cyclical or stationary on average, locking the camera, fixing the light, and, where the tool supports it, using the same image as first and last frame. A background loop also has a job: it must not compete with the person, headline or music in front of it.
</context>

<task>
Write a loop prompt for "[SCENE]", 8 seconds, used as a stream-background.

1. **Loop strategy.** Choose and explain the method that fits the scene:
   - **cyclical motion:** elements that repeat a whole number of times within 8 seconds (a pendulum, a rotating object, waves on a period that divides the loop length);
   - **stationary texture:** many small elements in constant, statistically even motion (rain, snow, flicker, particles, slow noise) where no single element must return;
   - **matched first and last frame:** generate from a start image and set the same image as the end frame if the tool supports it;
   - **crossfade fallback:** generate longer than needed and dissolve the tail into the head in the editor;
   - **ping-pong:** play forward then reversed; only for motion that looks natural backwards (not falling water, smoke or walking).
   If the scene requires one-way travel (a car driving through, a sunrise), say it cannot loop cleanly and propose the nearest loopable version.
2. **Prompt.** One paragraph: locked-off static camera (or a perfectly circular move only if the strategy supports it); the moving elements with motion type, speed and how they repeat; what stays still; constant light with no change in time of day or exposure; and the requirement that the scene at the end matches the beginning. Phrase it positively.
3. **Use-specific design.**
   - stream-background: low contrast and calm motion behind the streamer; keep a clear area for the camera box and alerts; avoid motion near the edges where overlays sit.
   - website-hero: very slow, subtle motion; a quiet area for the headline with enough contrast for text; keep file size small (short loop, modest resolution); provide a still poster frame for reduced-motion settings and slow connections.
   - visualiser: motion with a regular pulse; if the tempo is known, make the loop a whole number of bars (seconds per bar = 240 / BPM in 4/4) and say what it is.
4. **Settings.** Aspect ratio for the use, a low motion setting if the tool offers one, seed reuse for retries. Tell the user to check their tool's current first-frame and last-frame options.
5. **Seam check.** A checklist to run in an editor or player on loop: position of every visible element at the cut, brightness and colour at the cut, motion speed across the cut, any element that appears or vanishes, and viewing the loop at least five times in a row.
6. **Fallbacks.** Three fixes if the seam pops, each changing one thing.
7. Before answering, confirm that the prompt contains a static or circular camera, constant light, and a stated repeat or stationary pattern that fits 8 seconds.
</task>

<constraints>
- If no scene is given, or it is too vague to know what moves in the loop, ask for it in one question and stop.
- No rapid flashing: fewer than three flashes a second, and no large high-contrast flicker, to protect viewers with photosensitive conditions.
- No readable text or logos in the generated loop; add them as overlays.
- No tool, model or version names.
</constraints>

<output_format>
## Loop strategy
## Prompt
Code block.
## Settings
## Seam check
Checklist.
## Fallbacks
</output_format>
````

---

<a id="write-video-generation-prompt"></a>

## Write a video-generation prompt

`write-video-generation-prompt` · prompt · Video generation · https://hermes-ide.com/prompts/write-video-generation-prompt

Writes prompts for AI video generators covering subject, action, camera movement, lighting, style, audio and consistency, with a shot-by-shot version for longer clips. Use with text-to-video tools.

````markdown
<context>
Video models read a prompt as a description of one continuous shot. They do best with one clear subject, one main action described with a precise verb, a specified camera behaviour, and a defined look. They struggle with several simultaneous complex actions, crowds interacting, fast hand movements, legible text, cause-and-effect physics and keeping a character identical across separate generations. Clips are short (often 5 to 10 seconds per generation, depending on the tool), so longer pieces are built from shots, and consistency comes from repeating the exact same descriptions of characters, wardrobe, setting and style in every shot and, where the tool supports it, from reference images or image-to-video starting frames. Some current tools also generate synchronised audio (dialogue, effects, ambience) from the prompt.
</context>

<task>
Write a video-generation prompt for this idea.

<idea>
[IDEA]
</idea>

Target tool: generic
Total duration: 8 seconds

1. **Interpretation:** two or three lines on the video you are aiming for, the choices you made where the idea was open, and whether it fits in one shot. If the idea is too open to film (for example "something epic for my brand"), ask up to three questions (subject, setting, purpose or mood) and stop.
2. **Prompt** (one shot): write a single paragraph in this order, with concrete visual words:
   - **Shot and camera:** shot size, angle, lens feel, and one camera movement (static, slow push in, pull out, pan, tilt, tracking alongside, orbit, crane up, handheld), with its speed;
   - **Subject:** who or what, with fixed visual identifiers (age range, build, hair, clothing, colours);
   - **Action:** one main action with a precise verb and its pace, beginning to end within the shot;
   - **Setting:** place, time of day, weather, background activity kept simple;
   - **Light:** source, direction, quality and colour;
   - **Style:** live-action cinematic, documentary, animation style described by technique, film stock or grade, frame rate feel (slow motion, real time);
   - **Audio** (only for tools that generate sound): ambience, effects, and any short line of dialogue in quotes with who says it.
3. **Shot list:** if 8 seconds is longer than one generation in generic (assume about 5 to 10 seconds per shot if unsure, and say so), split it into shots in a table: shot number, duration, shot size and camera, action, transition (cut, match cut, continuous), and a full standalone prompt for each shot that repeats the consistency block. If one shot suffices, write "Single shot".
4. **Consistency block:** a reusable paragraph describing each recurring character, outfit, setting and visual style in fixed wording to paste into every shot; recommend generating a reference image first and using image-to-video or the tool's reference or character feature if it has one.
5. **Settings:** aspect ratio for the use (16:9 for widescreen, 9:16 for vertical social, 1:1 or 4:5 for feeds), duration per clip, and any settings the tool commonly offers (motion strength, seed reuse for consistency, resolution). Note that settings and limits change between versions and tell the user to check their tool's documentation.
6. **Variations:** two alternative prompts that change one decision each (camera, light or style) and what each changes.
7. **Troubleshooting:** three fixes specific to this video for common failures (subject morphing, too much or too little motion, the camera ignoring instructions, warped hands or faces, unwanted text).
</task>

<constraints>
- Keep each shot prompt focused: one subject focus, one main action, one camera movement. Move extra actions into separate shots.
- Do not ask the model to render readable text in the frame; add text in editing instead and say so.
- Do not write prompts that depict real, identifiable people (including public figures) doing or saying things they did not do, sexualised content of real people or any minors, or footage designed to pass as real news, evidence or a real brand's advertising.
- Do not imitate copyrighted characters or a living director's signature style by name; describe the visual qualities instead.
- Phrase prompts positively ("an empty street") rather than relying on negatives, unless the tool has a separate negative prompt field.
</constraints>

<output_format>
## Interpretation
## Prompt
In a code block, ready to paste.
## Shot list
Table, then one code block per shot; or "Single shot".
## Consistency block
In a code block.
## Settings
## Variations
## Troubleshooting
</output_format>
````

---

<a id="write-logo-sting-prompt"></a>

## Write an animated logo sting prompt

`write-logo-sting-prompt` · prompt · Video generation · https://hermes-ide.com/prompts/write-logo-sting-prompt

Writes prompts and a timing plan for a short animated logo sting, with a motion concept, frame-accurate timing, a sound cue and an end frame that keeps the real logo geometry exact.

````markdown
<context>
A logo sting is a few seconds of motion that ends on the logo, used at the start or end of videos, ads and streams. Video generators cannot be trusted with the logo itself: they redraw letterforms, bend proportions and shift colours, which no brand can accept on its final frame. The reliable approach is to generate the motion around the logo (light sweeps, particles, ink, paper, liquid, a background move) and composite the real vector logo on top, revealed by masks or matched to the generated motion, so the end frame is pixel-exact. The motion should express one idea that fits the brand, and the timing should be exact to the frame because stings are cut against music and sound.
</context>

<task>
Plan a 3-second premium logo sting.

<logo>
[LOGO_DESCRIPTION]
</logo>

1. If the logo's shapes and colours are not described or attached, ask for them in one message and stop.
2. **Motion concepts.** Propose three concepts that grow out of the logo's own shapes or meaning (a circle that rolls in, strokes that draw on, a fold that opens), each in two lines with how the real logo is revealed. Recommend one for premium and say why.
3. **Timing sheet.** For the recommended concept, at 24 and at 30 frames per second, break the 3 seconds into: anticipation, main motion, resolve into the logo, and a hold on the final logo of at least one second (or most of the length if the sting is very short). Give each phase in seconds and frames, and the frame where the sound hit lands.
4. **Generation prompts.** Write prompts only for elements a generator can make safely: the background plate, light sweeps, particles, textures, the transition elements. Each prompt states that the centre area stays clear for the logo, fixes the palette to the logo's colours, sets a static or simple camera, and gives duration and aspect ratio (16:9 by default, plus 9:16 and 1:1 versions if useful). Include one optional prompt for a motion reference (a rough of the movement) that a motion designer can follow, clearly marked as not for the final frame.
5. **Sound cue.** Describe the sound in words: type (whoosh, chime, thump, pluck, riser), sync points to the timing sheet, length, and tail. Keep it short enough that the last note decays during the hold. Point out that a separate jingle or sonic logo is a different deliverable.
6. **Compositing plan.** Steps in the editor or motion software: place the real vector logo, reveal it with a mask, wipe or opacity tied to the generated motion, match light and shadow, keep the final frame identical to the master logo, and export versions (with and without background, light and dark).
7. Before answering, check that no prompt asks the generator to draw the logo or its text, that phase timings add up to 3 seconds, and that the hold is long enough to read the logo.
</task>

<constraints>
- Never let generated pixels replace the real logo on the final frame.
- Do not imitate another company's sting, mascot or sonic logo.
- No tool, model or version names.
</constraints>

<output_format>
## Motion concepts
## Timing sheet
Table: Phase | Seconds | Frames at 24 | Frames at 30 | What happens.
## Generation prompts
One code block per element, each with a one-line purpose.
## Sound cue
## Compositing plan
Numbered steps.
</output_format>
````

---

<a id="write-image-to-video-motion-prompt"></a>

## Write an image-to-video motion prompt

`write-image-to-video-motion-prompt` · prompt · Video generation · https://hermes-ide.com/prompts/write-image-to-video-motion-prompt

Writes a motion prompt that animates a still image, naming what moves, what stays fixed, the camera move and the pace, so faces, text and logos do not warp.

````markdown
<context>
In image-to-video, the image already fixes the subject, composition and look. The prompt's only job is motion: what moves, how far, how fast, and what the camera does. Most failed animations come from asking for too much: a face that turns and speaks, a camera that orbits a product with a printed label, several things happening at once. The model then invents pixels it cannot see and faces melt, text scrambles and logos bend. A good motion prompt spends a small motion budget on a few elements, protects the fragile ones explicitly and describes the end state.
</context>

<task>
Write a motion prompt for this still.

<image>
[IMAGE]
</image>

<motion>
[MOTION]
</motion>

Length: 5 seconds. Camera: slow-push.

1. If the image is not attached and the description does not say what the subject is and where the faces, hands, text or logos sit, ask for those details in one message and stop.
2. **Frame map.** Sort what is in the frame into:
   - **Moves:** at most two or three elements, each with a type of motion (drift, sway, ripple, rise, blink, turn) and an amount (subtle, moderate, strong);
   - **Stays fixed:** the rest of the composition;
   - **Protected:** faces, hands, readable text, logos, product labels, fine patterns. For each, choose a protection: keep it still, keep motion away from it, or plan to restore it in the edit (overlay the original still or composite the logo afterwards).
   If the requested motion conflicts with a protected element (for example "she says hello" on a close-up face, or a full orbit around a labelled bottle), say so plainly and propose the nearest motion that will hold up.
3. **Motion prompt.** Write one paragraph in this order: camera behaviour (slow-push) with direction and speed over 5 seconds; the moving elements with precise verbs, amounts and pace; an explicit statement of what stays still; the end state of the shot. Do not re-describe the whole image; mention only what moves or must hold. Phrase it positively ("the bottle and its label stay still and sharp") rather than as a list of negatives, unless the tool has a separate negative field.
4. **Settings.** Recommend a motion amount (low, medium, high) if the tool exposes one, keep the source image's aspect ratio, and suggest reusing the seed across retries where possible. Tell the user to check their tool's current controls rather than assuming names.
5. **Gentler and bolder versions.** One version with half the motion and one with more, each with a line on the risk it trades.
6. **If it warps.** Three fixes specific to this image, each changing one thing only.
7. Before answering, check that the prompt moves no more than three elements, names every protected element as still or handled, and fits 5 seconds at the stated pace.
</task>

<constraints>
- Do not animate an image of a real, identifiable person to make them say or do things, or any image of a minor in a sexualised way; decline and offer a non-identifying alternative.
- Do not ask the model to change readable text or logos; restoring them in the edit is the reliable route.
- No tool, model or version names.
</constraints>

<output_format>
## Frame map
Three short lists: Moves, Stays fixed, Protected (with the protection chosen).
## Motion prompt
Code block.
## Settings
## Gentler and bolder versions
Two code blocks with a one-line note each.
## If it warps
Three numbered fixes.
</output_format>

<examples>
<example>
Image: a ceramic mug of coffee on a wooden table by a rainy window, the café's logo printed on the mug. Motion: "cosy, alive". Camera: slow-push.
Motion prompt: "Very slow push-in toward the mug over five seconds. Thin steam curls upward from the coffee and drifts left. Raindrops slide down the window glass behind. The mug, its printed logo and the table stay completely still and sharp. Ends on a slightly closer framing of the mug with steam still rising."
</example>
</examples>
````

---

<a id="write-b-roll-prompts"></a>

## Write b-roll generation prompts

`write-b-roll-prompts` · prompt · Video generation · https://hermes-ide.com/prompts/write-b-roll-prompts

Writes generated b-roll prompts matched to a video's script beats and existing footage look, with lens, grade and motion cues so clips blend in, and flags beats that need real footage.

````markdown
<context>
B-roll illustrates what the speaker is saying, covers cuts and keeps the eye moving. Generated b-roll stands out when its look does not match the A-roll: a glossy, perfectly lit, slow-motion clip dropped into handheld phone footage reads as fake at once. It also creates an honesty problem when it stands in for something the viewer will take as a record of real events: a news scene, a real place, a customer, a product test. Good generated b-roll copies the real footage's camera and grade, illustrates ideas and generic actions, and leaves documentary moments to real footage.
</context>

<task>
Write up to 8 b-roll clip prompts for these script beats.

<script_beats>
[SCRIPT_BEATS]
</script_beats>

<footage_look>
[FOOTAGE_LOOK]
</footage_look>

1. If the footage look is too vague to match (for example "normal"), ask for camera, frame rate, grade and light in one message, or for a frame from the footage, and stop.
2. **Look match block.** Turn the footage look into a reusable block: lens and field of view, sensor feel (phone, cinema camera), frame rate and motion feel (real-time, no slow motion unless the footage uses it), camera support (handheld micro-shake, tripod), colour temperature and grade, contrast, grain or sharpening, typical light source.
3. **B-roll plan.** For each beat, decide what picture best supports the line: a concrete action, a detail, a place or a metaphor. Prefer specific, ordinary images over stock clichés. Size each clip to the beat's length. If there are more beats than 8, prioritise beats where the speaker is static longest or where an edit needs covering.
4. **Honesty check.** Mark any beat where generated footage would mislead because the viewer would take it as real: news or historic events, a named real place shown as it is now, real people or customers, testimonials, product results or tests, data shown on a screen, anything presented as evidence. For those, recommend real footage, licensed stock with accurate captions, or a clearly stylised illustration instead.
5. **Clip prompts.** For each clip: the look match block pasted verbatim, shot size and angle, one action, setting, light consistent with the A-roll, duration. No readable text, screens with legible content or logos.
6. Before answering, check that every prompt includes the look block unchanged and that no flagged beat received a realistic generated clip.
</task>

<constraints>
- No real, identifiable people and no real brands' products or logos in generated clips.
- Do not suggest labelling generated clips as real footage. Suggest disclosure where the platform or the content calls for it, and tell the user to check their platform's current rules on synthetic media.
- No tool, model or version names.
</constraints>

<output_format>
## Look match block
Code block.
## B-roll plan
Table: # | Beat or timestamp | Line | Picture | Seconds | Generated or real.
## Clip prompts
One code block per generated clip.
## Use real footage here
The flagged beats, each with why and what to use instead.
## Blending tips
Three to five tips for matching in the edit (grain, colour match, frame rate conform, speed, crop).
</output_format>
````

---

<a id="write-explainer-animation-prompts"></a>

## Write explainer animation scene prompts

`write-explainer-animation-prompts` · prompt · Video generation · https://hermes-ide.com/prompts/write-explainer-animation-prompts

Turns an explainer voice-over script into timed motion-graphics scene prompts that share one visual system and match each scene to the line it illustrates.

````markdown
<context>
An explainer works when every picture does one job: it shows the idea the narrator is saying at that moment, in a visual language the viewer learns once and can then read without effort. Generated explainer scenes usually fail in three ways: each scene invents a new style, the visuals are literal stock-photo clichés (a lightbulb for "idea", handshakes for "partnership") or decoration that ignores the line, and the timing does not match the narration so pictures lag behind words. Text in generated frames also comes out garbled, so labels and captions belong in the edit.
</context>

<task>
Build scene prompts for this voice-over in the `flat-2d` style, for a 60-second video.

<script>
[SCRIPT]
</script>

1. **Timing check.** Count the words. Estimate narration time at about 150 words per minute (2.5 words per second) for a calm explainer. Compare with 60 seconds. If the script runs more than 15% over or under, say by how much and suggest which lines to cut or where to hold visuals; do not rewrite the script yourself. If the script is not a voice-over (for example bullet notes), ask for the finished narration and stop.
2. **Visual system.** Define once, for reuse in every prompt:
   - palette (4 to 6 colours by plain name, with one accent kept for the key idea);
   - shape language and line weight;
   - a recurring character or icon set, described exactly, if the script has a protagonist (a customer, a cell, a parcel);
   - backgrounds (plain, gradient or simple set);
   - motion grammar (how things enter, transform and leave; typical pace; the transition used between scenes);
   - camera (mostly static or slow push, appropriate to flat-2d).
3. **Scene breakdown.** Give each voice-over line, or each pair of short lines, one scene. For each, choose a visual metaphor that shows the mechanism, not a cliché: a queue shrinking for "faster checkout", a parcel route redrawing itself for "rerouted delivery". For abstract lines, offer the metaphor and one alternative. Assign a time range from your word count so the visual lands with its words; the scene times must add up to the narration time.
4. **Scene prompts.** For each scene: the visual system tokens, the elements on screen, the one main motion with start and end state, the camera, the transition in, and the duration. Keep each scene to one idea and no more than three moving elements.
5. **Edit notes.** List every label, number, caption or logo that should be added as text in the editor rather than generated, with its scene and time. Add accessibility notes: captions for the whole voice-over, sufficient colour contrast for any on-screen text, no rapid flashing.
6. Before answering, check: every script line maps to a scene; times add up; every prompt repeats the same palette and style tokens; no prompt asks for readable text.
</task>

<constraints>
- Do not change, add or drop words in the script. If a line is unclear or a claim looks wrong, flag it in the timing check for the author.
- No real people or real brands' logos and characters in the visuals unless the user owns them and says so; logos go in the edit.
- No tool, model or version names.
</constraints>

<output_format>
## Timing check
Word count, estimated seconds, difference from 60, and any suggestions.
## Visual system
Code block with the reusable tokens.
## Scene table
| # | Time | Voice-over line | Visual idea | Motion | Transition |
## Scene prompts
Numbered code blocks.
## Edit notes
Text overlays and accessibility.
</output_format>
````

---

<a id="write-music-video-shot-prompts"></a>

## Write music video shot prompts

`write-music-video-shot-prompts` · prompt · Video generation · https://hermes-ide.com/prompts/write-music-video-shot-prompts

Plans a beat-synced music video for the creator's own song, mapping each section to visuals with a recurring motif, cut points on the beat and a generation prompt for every shot.

````markdown
<context>
A music video lands when picture and song move together: cuts fall on beats and bar lines, the visual energy rises and falls with the arrangement, and one image keeps returning and changing so the video feels authored rather than assembled. Generated clips are short and independent, which suits fast cutting but tempts people to string together unrelated pretty shots. Planning the beat grid first, then the sections, then the motif, keeps the video coherent and makes every clip a deliberate length.
</context>

<task>
Plan a music video for this song. Mood: [MOOD].

<song_structure>
[SONG_STRUCTURE]
</song_structure>

1. This prompt is for the creator's own music. If the song is someone else's, or section timings are missing, ask in one message (for the creator's own track, or for timings) and stop. If the BPM is missing but timings are given, estimate it from the section lengths and say it is an estimate.
2. **Concept.** If a concept was given, restate it in two lines and name the visual motif. If not, propose three distinct concepts (for example performance, narrative, abstract), each with a motif, and pick the one that best fits [MOOD], saying why.
3. **Beat grid.** From the BPM, give seconds per beat (60 / BPM) and per bar (4 beats in 4/4, or as the song's meter), and the cut lengths you will use: for example 1 bar for verses, half a bar in the chorus, 2 bars for the intro. Round clip lengths to whole beats.
4. **Section map.** For each section: time range, energy level (1 to 5), what the visuals do (location, action, palette shift), how the motif appears and evolves, and the cut rhythm. Energy and cut rate should rise into choruses and drop for breakdowns; the final chorus or outro should pay off the motif.
5. **Shot list.** Number every shot with section, start time, length in beats and seconds, shot size and camera, action, and which beat the cut lands on (downbeat, snare, the first beat of the bar). Flag the hero shots worth extra takes.
6. **Shot prompts.** One standalone prompt per shot with a fixed look block and fixed character blocks (if any) pasted word for word, one action, camera behaviour and duration.
7. Before answering, check that shot lengths in each section add up to the section's duration within a beat, that the motif appears in at least the intro, every chorus and the ending, and that every prompt reuses the same look block.
</task>

<constraints>
- Do not reproduce lyrics the creator did not paste, and do not quote lyrics from other songs.
- Characters are original or the creator themself (performance shots to be filmed or generated from the creator's own reference with their consent). No real celebrities or other artists' likenesses.
- No readable text generated in frame; titles and lyric captions go in the edit.
- Avoid rapid full-screen flashing (more than three flashes a second) and warn if the concept depends on strobing.
- No tool, model or version names.
</constraints>

<output_format>
## Concept
## Beat grid
## Section map
Table: Section | Time | Energy | Visuals | Motif | Cut rhythm.
## Shot list
Table: # | Section | Start | Beats / seconds | Shot and camera | Action | Cut on.
## Shot prompts
Look block and character blocks in one code block, then one code block per shot.
## Edit notes
Sync tips, where to use speed ramps or holds, and captions.
</output_format>
````

---

<a id="write-product-video-prompt"></a>

## Write product video prompts

`write-product-video-prompt` · prompt · Video generation · https://hermes-ide.com/prompts/write-product-video-prompt

Writes video-generation prompts for product reveals, spins, detail shots and in-use moments that keep shape, colour and label accurate, sized for listings, social ads or a site hero.

````markdown
<context>
Product video is the least forgiving use of video generation. A buyer compares the video with the box that arrives: if the bottle was taller on screen, the green a different green, the label rearranged or a feature shown that the product does not have, the result is returns, complaints and, for ads and listings, possible breaches of advertising and marketplace rules. Generators also redraw labels and logos as near-miss shapes. The workable approach is to describe the product with fixed, measurable wording, keep motion modest around it, get printed graphics from the real packshot in the edit, and never show a capability that has not been confirmed.
</context>

<task>
Plan 4 product video shots for social-ad.

<product>
[PRODUCT_DESCRIPTION]
</product>

1. If the description lacks the product's shape and proportions, its exact colours, or where the label or logo sits, ask for them (or for reference photos) in one message and stop. If 4 is outside 2 to 8, use the nearest bound and say so.
2. **Product truth sheet.** One fixed block to paste into every prompt: object type, proportions (for example "height about twice the width"), materials and finish (matte, gloss, brushed), colours by plain name plus any code given, label or logo position and size, parts that move and how. Then a list of **claims allowed** (only what the user supplied) and **never show** (features, results or scale not supplied).
3. **Shot plan.** Choose from these shot types to fit social-ad:
   - reveal (the product enters frame, a cover lifts, light sweeps on);
   - turn (a partial turntable rotation, usually 30 to 90 degrees, rather than a full spin, so the model does not invent the back);
   - detail (macro on texture, a seam, a button);
   - in use (a hand or person using it as it is really used, with the real result);
   - end frame (a clean packshot for the call to action).
   For **listing**: neutral background, true colour, no exaggerated effects, realistic scale with a familiar object or hand. For **social-ad**: 9:16, the product or the problem it solves on screen in the first two seconds, a clear end frame. For **website-hero**: 16:9 or wider, slow continuous motion, room for headline text, a loop-friendly ending.
4. **Shot prompts.** For each shot: shot size and angle, camera move and speed, the truth sheet pasted verbatim, the single action, setting and surface, lighting (key direction, softness, reflections controlled for glossy surfaces), duration, aspect ratio. Keep hands simple (one hand, a clear grip) and away from the label.
5. **Label and logo plan.** Recommend generating with the label area plain or turned slightly away, then tracking or compositing the real label, logo and any on-screen text from the packshot in the edit. Name the shots where that matters most.
6. **Accuracy check.** Before answering, compare every prompt against the truth sheet and the allowed claims: same proportions, colours and label position in every shot; no feature, result, size or ingredient that was not supplied. Report what you checked and anything you removed.
</task>

<constraints>
- Do not depict results, performance or features the user did not state (a blender crushing ice, a cream removing wrinkles, waterproofing). If the user asks for one that is not in the product description, ask them to confirm it is true.
- Do not imitate a competitor's branding, a real celebrity or a real influencer.
- Remind the user that platforms and regulators may require generated or altered product imagery to stay accurate, and that some listing sites restrict generated media; tell them to check the current rules for their channel.
- No tool, model or version names.
</constraints>

<output_format>
## Product truth sheet
Code block, then "Claims allowed" and "Never show" lists.
## Shot plan
Table: # | Type | Seconds | Shot and camera | Action | Purpose.
## Shot prompts
One code block per shot.
## Label and logo plan
## Accuracy check
Bulleted results.
</output_format>
````