Achref23illi/knox-prompts

Knox Prompts — image & video prompt engine

Purpose

Turn ideas into production-ready image and video prompts: Subject → Environment → Camera → Lighting → Mood → Style, plus depth and framing, realism, color grading, and targeted negatives.

Prompting is direction, not description. The less direction you give, the more the model fills in blanks — and when a model fills in blanks it picks the most average version of the idea. This skill's job is to remove the blanks.

Scope. This skill owns how the frame looks and moves. It does not own what a person says, which ad angle a campaign uses, or whether a batch of deliverables is on track. If the user has a separate campaign, script, or character-identity skill, invoke that one first and come back here for the shot prompts.

---

The 6-part framework (reason through all six, every time)

1. Subject — who/what; age, hair, clothing, expression, posture, exact action, camera awareness, gaze
2. Environment — specific place, time, weather, lived-in objects, textures, atmosphere, busyness
3. Camera — shot size, angle, position (height / offset / distance), lens mm, aperture/depth, device or body, lens character, focus target
4. Lighting — physical source, direction, quality, fill (only if it could exist), shadows/contrast, white balance, exposure. Never just "cinematic lighting"
5. Mood — 1–2 words, maximum
6. Style — optional; only if it helps. Over-styling kills realism

Plus depth & framing: foreground / midground / background, atmosphere (haze, dust, fog, smoke — only if physically plausible), rule of thirds, leading lines, lead room.

Simple recipe: [Subject] in [environment], [camera], [lighting source + direction], [mood], [style]

Whenever a prompt is not working, one of those six parts is missing, weak, or contradicting another.

---

Final prompt assembly order (strict)

1. Subject — identity anchor + action + expression + posture + outfit priority
2. Environment
3. Depth & composition
4. Camera
5. Lighting
6. Mood
7. Realism details
8. Style (if any)
9. Color grade — one palette, must match the lighting, placed near the end (models weight the end)
10. Negative constraints — only the failure modes actually likely for this shot

---

Realism standard

Preserve: visible pores, uneven skin tone, under-eye texture, natural sheen, individual beard and hair strands, asymmetry, flyaways, fabric weight and wrinkles, real reflections, plausible shadows, scene-appropriate grain, noise, focus miss, motion blur, imperfect framing.

Pick 1–3 imperfections that fit this shot — not all of them.

Avoid these words: cinematic masterpiece, hyper-detailed, ultra-glossy, perfect lighting, 8K perfection, "colorful".

Avoid these looks: plastic skin, beauty filter, fake rim light, studio fill in a casual scene, HDR-everything, fake bokeh, CGI feel, impossible light, dead-center-by-default framing, stock-photo posing (a "running man" pose is not a sprint).

Capture context boosts realism. Say who is holding the camera, how the device behaves (auto exposure, HDR, no flash, 1× lens), whether the moment was planned or caught, and what happened 2–5 seconds before or after.

---

Camera vocabulary

Angles — eye level, three-quarter eye-level, frontal, profile, back shot, over-the-shoulder, low angle, worm's-eye, high angle, Dutch, overhead / top-down, POV.

Shot sizes — extreme wide, wide, suspense wide, full, medium-wide, medium, medium close-up, close-up, extreme close-up.

Lenses — 16–24mm wide, 24–28mm phone/documentary, 35mm street, 50mm normal, 85mm portrait, 200mm telephoto.

Aperture — f/1.4 melts the background (only worth asking for if the background is actually far away); f/8 keeps everything sharp. A phone is an ordinary 1× lens: no fake bokeh.

Bodies — ARRI Alexa, Sony Venice 2, RED V-Raptor, Hasselblad medium format, iPhone 14/15/16 Pro (front 24–26mm, rear 24mm).

Lens character — clean modern, warm organic (Cooke), vintage, anamorphic.

---

Color grading map

Mood	Palette
Epic / heroic	warm golds, deep blue shadows
Dark / moody	desaturated, teal-orange
Romantic	soft pink and gold, haze
Vintage	faded warm, lifted blacks, 35mm grain
Clean / corporate	bright, slightly cool, high-key
Documentary	rich greens, warm earth
Melancholy	blue-grey, muted
Action	high contrast, punchy
Luxury	deep blacks, gold accents

Film references that models read well: Blade Runner 2049, Mad Max, The Matrix, Wes Anderson, Ozark / Gone Girl, La La Land, John Wick, Dune, Barbie, The Revenant.

One palette only. Place it at the END. It must be compatible with the lighting you described.

---

Modes

1. New prompt (default) Ask only for the missing essentials: the idea (who + doing what + where + why), platform and aspect ratio, what the viewer should notice first, what must not happen. If the user says "you decide", decide. Reason through the six parts plus depth. Choose camera and light yourself if unspecified. Output: (a) a 2–4 line creative recommendation, only if a real decision was made; (b) one copy-ready final prompt in assembly order; (c) a short hard-constraints block; (d) the exact attachment list with file paths and order.

2. Full brief — a planned post or campaign. Walk the brief questions in order (subject, goal, platform, audience, place, time, wardrobe, props, camera, light, mood, grade, constraints), then assemble.

3. Improve / fix fake — the user pastes a prompt or says the result looks AI. Diagnose which of the six parts is weak, or which giveaway is present. Show only: cause → fix → rewritten prompt.

4. Edit prompt — change one variable (outfit, lighting, camera, background). Change only that layer; keep everything else verbatim.

5. Critique — evaluate against the framework: identity vs. the master reference (give a % and name the drifted features), subject clarity, environment, camera, light plausibility, depth, composition, realism, color, AI giveaways, continuity with adjacent shots. Say what works. Prefer a targeted edit over a regeneration. For video, extract a frame strip and a motion/luma curve with ffmpeg and report where the action actually lands.

6. Video prompt — see below.

7. JSON — fields: subject, environment, composition, camera, lighting, mood, realism, style, color_grade, negative, final_prompt. Video adds camera_move, subject_motion, environment_motion, duration, dialogue, sfx.

8. Variations — N versions. Change angle / mood / setting / hook / style; keep the core.

9. Style or grade lookup — pull from the tables above; style keywords go at the END of the prompt.

10. Shot bible — multi-shot video. Per shot: duration, technique, image prompt, video prompt, attachment list. Plus a global identity line, the grade, the cut list, and a continuity-reference folder. Save to projects/<name>/SHOT_BIBLE.md.

---

Image-model prompt rules (hard-won)

• Never write @name inside an image or video prompt. @character tokens are assistant routing only — the model cannot resolve them. Spell the identity out: "the man in the first attached photo, exact same face: [3–5 traits]".
• Number the attachments in the prompt ("first attached photo", "second…") and say which face maps to which person. Tell the user the exact attach order and the file paths.
• Two or more characters in one frame = drift for both. Make one the main subject (gets the light, the focus, and the full identity description) and describe the other in text only, "generic, secondary". Or do two passes: generate the scene, then edit-mode per face with that character's master reference plus a "FACE IDENTITY RESTORATION ONLY" line.
• Never attach a filmstrip or contact sheet as a reference. Single frames only.
• Role-lock every reference. When more than one image is attached, say what each is allowed to control: "character sheet = identity only, product image = packaging only, tee image = wardrobe only". Unlabelled references bleed into each other.
• Continuity. Every new shot in a known location attaches a single frame pulled from the approved clip of that location, first in the attachment order, and the prompt names 4–5 concrete architectural details from it (building height, balconies, shutters, pavement material, parked vehicles). One adjective — "sandstone" — will build a different city every time.
• Crowds: ask for variety explicitly (heights, builds, genders, clothing) or you get clones.
• Screen text and logos: "keep exactly as attached, undistorted, correctly spelled" — and expect to mask it in the edit anyway.

---

Video prompting (Kling / Seedance / image-to-video)

Golden rules

1. Camera move FIRST. Open the prompt with the move — image-to-video models read it immediately.
2. The still must physically allow the move. Crane up → open sky above, nothing between camera and sky. Door shots → the door state must make the direction unambiguous (a closed door means push out; an open door is ambiguous). Sliding doors animate better than hinged doors. Orbit → a background with real depth. Over-the-shoulder → a character present.
3. One shot = one job. Camera move OR blocking OR dialogue. Stacking two gives you identity drift and invented people.
4. Big camera moves → start frame + end frame. Generate the end frame from the start frame: "the exact same scene, camera has risen 3 m / moved back 4 m, everything else unchanged". Then run the model in start+end mode.
5. Motivate the move. The camera moves at a trigger. Unmotivated movement drifts.
6. Ease-in, ease-out, locked final frame. Say it out loud.
7. Safer version for AI: a 90–180° arc, not 360; one or two foreground layers; one reveal; a short push-in, not a long float.
8. Say the failure before it happens: "stays flat and rigid", "does not enter the store", "no new people", "no camera pass-through", "the people keep their shape", "moves forward and left only".
9. Subjects who shouldn't act: say so. "Stay perfectly still, backs to camera, unchanged." The model invents action if you don't forbid it.
10. Realism motion: micro handheld shake, autofocus hesitation, exposure breathing, a blink, a breath, fabric shift. Subtle.
11. Identity line (no @): "the man's face stays identical to the start frame throughout, no drift, no morphing."
12. Models front-load stillness and act in the last 1.5–2 s. Plan your trims from the tail: generate 5 s, cut 2–3 s. Static-object shots need a micro-move or you accept one dead beat.
13. They can't spell. Screen text and caller names get masked or overlaid in the edit.
14. Model routing: people interacting, blocking, or turning → Seedance. Camera moves, objects, POV, vehicles, reveals → Kling.
15. Dolly zoom rarely lands. You'll get a push-in. Accept it, or use start+end frames.
16. Echo-print / step-print: describe it ("ghost trails that lag a fraction of a second and snap back, accordion blur"). If it's ignored, do it in the edit: duplicate layer, offset 3 frames, 40% opacity, ×3.
17. Use amplitude words for small motion: "rattles hard, jumps, flashes then dims" — not "vibrates gently".

Single-shot structure

[Camera move + speed] toward/around/b

…(truncated)…