prompt-craft

Imported from petaverse-cloud/pv-skilyst-agent (skills/official/prompt-craft/SKILL.md). Install upstream with npx skills add petaverse-cloud/pv-skilyst-agent --skill prompt-craft. Copyright stays with the author.

name: prompt-craft description: Camera-and-prompt craft for the video nodes on the Skilyst/Beehive platform — the cross-model structural skeleton, the per-model deltas that actually change the output, the five-part reference-lock syntax that keeps a character consistent, and the dialogue pacing rule. Use when writing or reviewing any generation prompt, before submitting a job; contains no node dependency of its own. license: MIT compatibility: Requires nothing but the skilyst runtime (>=0.1.0): pure methodology, no network, no credential. It declares requires.nodes = [] on purpose so it can never make a platform tool available. metadata: skilyst.skill_id: skilyst/prompt-craft skilyst.version: "1.0.0"

Prompt craft: the four rules that survive contact with a model

Load this before writing or reviewing a generation prompt. It carries only the rules that change output; the full keyword tables, the model-by-model formula comparison and the category templates live in the lane research corpus (prompt-library.md, cinematic-language-guide.md, serialized-drama-guide.md under research/visual-language/ — read them when a shot needs depth, do not duplicate them here).

Rule 1 — the skeleton (slot order is the technique)

[shot size] + [subject + appearance anchors] + [ONE primary action + micro-action]
+ [scene, 3-5 elements] + [light] + [style/mood] + [audio, if the model has it]
+ [negative / constraints]

• Subject must be trackable: age, hair, wardrobe colour, accessories, build. Visual anchors are what lets the model recognise the same person across shots.
• One primary action per shot. A 5-10 s shot cannot carry a second action line; use environmental micro-motion (breath, wind, steam, background traffic) to add life — but never let it compete with the subject's action.
• Camera at the head or in its own sentence. Camera commands are positional, not decorative.
• Light must be drawable: golden hour backlight with warm rim illumination, never good lighting.
• Negatives are model-dependent (see Rule 2) and are only one of the levers.
• Temporal phrasing buys rhythm: begins with … then … finally …, or hard segmentation where the model supports it (0-3s: … 3-6s: …).

Rule 2 — the per-model deltas (this is the part that changes output)

Cross-model calibration worth keeping: subject + movement + scene are the required triple everywhere; framing/light/mood are the gain layer. Models are weak on counts (a stated number of objects is usually wrong) and on complex physics (bounces, parabolas); safety filters false-positive on innocent words, so describe clothing instead of bodies.

Rule 3 — the five-part reference lock (when a character must repeat)

Handing over a picture is not a lock: the model reads an unexplained image as a style reference. Write all five parts:

1. Token declaration — name each reference (@image1, @location1, @hands1) and reuse the identical token throughout.
2. Inheritance list, enumerated — face, face shape, skin tone, hairstyle, height, build, each wardrobe item by name, footwear, bearing. Name each garment separately; "the same outfit" inherits nothing.
3. Non-inheritance declaration — background, furniture, pose, camera angle, original lighting, any text: state that they do NOT carry over. Skip this and the reference image's composition migrates into the shot.
4. Cross-shot restatement — on turns, occlusions and fast motion, restate that it is the same face.
5. Negatives — no cloning/duplication, no averaged faces, no swapping attributes between characters.

Sketch or line-art references additionally need render as realistic live-action.

Rule 4 — dialogue and pacing

• ≤8 words per 3 seconds. A 15-word line inside 3.3 s came in audibly too fast in the lane's own measurement: cut the line, never speed the render up.
• Declare the language on its own line, keep the line in the script the model was told to speak, and state no silent moments and no voice-over for continuous dialogue — otherwise the model scores the scene and animates the mouth anyway.
• Write reactions as a chain, not a list: hears → pauses → expression changes → body follows → aftermath. Emotional state is written as behaviour (looks away, pressed-together lips, slower delivery), not as a symptom.
• Move proper nouns, numbers and names out of the spoken line and onto the screen in post: they are mispronounced and they eat the pacing budget.
• One speaker per shot; two speakers split the mouth budget.

Narrative craft worth applying at sequence level

• Title each act — the title is what bounds how much information fits.
• Write the reversal as a picture (the cloud layer parts over the empty city) instead of announcing a reveal.
• Bind the emotional turn to a light change (golden hour → blue hour).
• Over 30 s, split into separate prompts and cut them together — every video node here has a hard per-job cap; the splice craft is in skilyst/embed-video.

How to use this skill

1. Decide the shot's single intent → pick the shot size (see Rule 1).
2. Write the skeleton, then adapt to the target node's row in Rule 2.
3. If a character repeats, write the five-part lock (Rule 3); for dialogue shots, apply Rule 4 and then hand off to skilyst/lipsync-audio-refs.
4. Verify the field names and value ranges against that node's own input_schema (GET /api/v1/nodes/<node_id>) before submitting — prompt craft decides the content, input_schema decides whether the job is accepted.

Deeper material (keyword slots by category, the 21-source formula comparison, category templates) is in the lane corpus referenced above.