Prompt Quality Checklist

intent

this skill audits AI image and video prompts before generation by running 10 systematic checks on each shot description. use it when building short-form video storyboards, batch-generating assets, reviewing teammate prompts, or doing weekly error retrospectives. the goal is to catch vague language, spatial inconsistencies, lighting contradictions, and missing asset anchors that cause generation misalignment across multiple shots. run the checklist before hitting generate. every check must pass or the prompt goes back for revision.

inputs

required context:

• raw prompt text (english or chinese, any length)
• optional: character reference image (PNG/JPG, for shot consistency checks)
• optional: location/venue reference image (for scene anchoring)
• optional: existing shot list or storyboard document (for cross-shot consistency)

external connections:

• image generation tool (Runway, Pika, LibLib, 即梦, 海螺, or equivalent). no auth required for checklist itself; used for prompt submission after passing review.
• optional asset management system (file paths or database URLs for character_ref, location_ref, prop_ref, costume_ref). if you use one, store asset paths in standardized [ASSET] format at prompt end.

no env vars or API keys needed for the checklist itself. this is a synchronous review process.

procedure

follow this sequence. do not skip items. mark each ✅ pass or ❌ fail.

step 1: time of day marker

• input: raw prompt text
• check: does the first sentence or clause explicitly state a time (day, night, dawn, dusk, midday, golden hour, 白天, 深夜, 黄昏, 正午, etc.)?
• output: marked ✅ if present, ❌ if absent or vague (e.g., "natural lighting" does not count). reason: AI light behavior changes drastically between day/night. omitting time delegates mood to randomness.
• example fail: "a woman in red standing on a rooftop"
• example pass: "night, a woman in red standing on a rooftop, city lights below"

step 2: character/subject description

• input: prompt + any character reference image provided
• check: is the person's appearance, clothing, pose, hair, and distinguishing details described in specific terms, not generic adjectives (beautiful, lovely, attractive)? does the description lock to a reference asset or costume code?
• output: marked ✅ if description includes ≥3 concrete details (fabric type, color, fit, accessory, pose) + asset path or reference label. marked ❌ if vague or contradicts the reference image. reason: "pretty girl" triggers different results per model. specific + asset path = consistency across shots.
• example fail: "a beautiful girl wearing a white dress"
• example pass: "character ref char_001.png, white silk qipao, hair in bun, right hand touching earring, same appearance as costume sheet"

step 3: scene/space anchor

• input: prompt scene description + optional location reference
• check: are there ≥3 specific hard props, architecture elements, or locations named (desk, bookshelf, window, wall color, floor type, specific venue)? avoid generic labels like "office" or "street."
• output: marked ✅ if ≥3 named, concrete objects with spatial relation (left wall, right corner, center desk). marked ❌ if fuzzy ("a room," "an interior," "luxury office"). reason: "nice office" diverges wildly across models. hard anchors (mahogany desk, green banker lamp, venetian blinds) pin the scene.
• example fail: "interior of a luxury office"
• example pass: "interior office_int_003, dark walnut desk center, green banker lamp on left, venetian blind window on right wall"

step 4: light source direction

• input: prompt lighting description
• check: is light direction specified with explicit angle or position (left, right, above, 45-degree, overhead, rim light, backlight)? are vague terms like "natural lighting," "柔和光线," "soft," or "cinematic" completely absent?
• output: marked ✅ if direction is concrete and quantified (e.g., "single light from left at 45 degrees, hard rim on right shoulder"). marked ❌ if vague or defaulting to "natural." reason: "soft lighting" is undefined. same scene with side light, back light, or top light produces utterly different moods and spatial reads.
• example fail: "soft natural lighting, cinematic"
• example pass: "single light source from left at 45 degrees, hard rim light on character right shoulder, no fill light"

step 5: ground contact shadow

• input: prompt and character placement description
• check: does the prompt explicitly mention a shadow on the ground or contact surface beneath the subject (shadow on floor, ground shadow, foot shadow, shadow stretching to right, etc.)?
• output: marked ✅ if ground shadow is described with direction and approximate length. marked ❌ if absent or only mentions character shadow without placement. reason: missing ground shadow makes subjects look composited/floating. ground contact is the baseline for spatial realism.
• example fail: "full body shot of a woman standing on the street"
• example pass: "full body shot, woman standing on wet asphalt, dark ground shadow directly beneath feet, shadow stretching 1.5m to right"

step 6: material/texture consistency

• input: prompt texture and finish language
• check: does the prompt use retouching/beauty terms (smooth skin, porcelain skin, flawless, matte, glossy, 磨皮, 柔焦)? are material descriptions contradictory (e.g., glossy lips + matte skin + rough fabric)?
• output: marked ✅ if materials are unified in finish and description avoids beauty mode language. marked ❌ if retouching vocab is present or materials conflict. reason: "porcelain skin" + "matte eyeshadow" triggers AI beautification mode, fracturing texture unity. stick to one material paradigm (realistic, stylized, weathered, etc.).
• example fail: "porcelain smooth skin, matte eyeshadow, glossy lips"
• example pass: "realistic skin texture with visible pores, matte eyeshadow, semi-matte lips, consistent material finish"

step 7: three-layer depth (foreground, midground, background)

• input: shot framing and composition description
• check: does the prompt define or imply objects or elements in three spatial layers? at minimum: midground (subject), foreground element (blurred or sharp), background element (environment/skyline/window)?
• output: marked ✅ if all three layers are named or clearly implied (e.g., "character in middle, bookshelf blurred left foreground, rain window background"). marked ❌ if flat or only subject + void. reason: even close-ups need Z-axis separation (eyelashes foreground, face midground, blurred background). depth prevents flatness.
• example fail: "close-up shot of a character's face"
• example pass: "medium shot, character in midground, blurred ancient bookshelf left foreground, rain-streaked window right background, shallow depth of field, focus on character"

step 8: shot type matches emotion/narrative

• input: shot type descriptor (wide, medium, close-up, extreme wide, low angle, high angle) + scene intent
• check: does the shot type serve the emotional or narrative goal of the moment? (e.g., isolation = wide + vast background, intimacy = close-up, power = low angle, overwhelm = extreme wide)
• output: marked ✅ if shot type aligns with stated mood or scene function. marked ❌ if contradictory or random. reason: shot type is a narrative tool, not decoration. wide shot of a tiny figure = loneliness. close-up of hands touching = intimacy.
• example fail: "extreme close-up of a sad woman with vast ocean behind her" (composition contradicts emotion, close-up negates vastness)
• example pass: "full body shot, woman standing alone on shore at dusk, tiny against vast ocean, vastness reinforces loneliness"

step 9: shadow projection direction consistency with light

• input: step 4 light direction + step 5 ground shadow direction
• check: do light angle and shadow angle point in opposite directions? (light left = shadow right; light 45-degree upper left = shadow 45-degree lower right; light top = short shadow below)
• output: marked ✅ if shadow direction is opposite and proportional to light angle. marked ❌ if shadow points same direction as light or absent. reason: shadow direction violates physics if it doesn't oppose light source. inconsistency breaks immersion and reads as AI error.
• example fail: "light from left at 45 degrees, shadow cast to left side of character"
• example pass: "light from left at 45 degrees, hard shadow stretching 2m to right of character, shadow angle mirrors light angle"

step 10: asset anchor annotation

• input: completed prompt + all reference images and prop/location/costume codes used
• check: is there a final [ASSET] block at the prompt end listing all reference paths or file codes in consistent format (char_ref, location_ref, prop_ref, costume_ref)?
• output: marked ✅ if [ASSET] block is present with all relevant paths in standardized format. marked ❌ if missing or inconsistent. reason: asset anchors enable shot-to-shot consistency and prompt reproducibility. without them, each generation is one-off; with them, you can audit and iterate.
• standard format: [ASSET] character: char_001.png | location: loc_venue_02.jpg | prop: sword_03.png | costume: outfit_01.jpg
• example fail: "a warrior holding a sword in dark forest" (no asset block)
• example pass: "... dark forest, warrior longsword, low angle\n[ASSET] character: warrior_m_001.png | location: forest_dark_001.png | prop: weapon_sword_002.png"

final step: tally and decision

• input: all 10 step results (✅ or ❌)
• output: if all 10 = ✅, mark prompt PASS and route to generation tool. if any = ❌, mark FAIL, log which steps failed, return prompt to author with specific fix notes.

decision points

if prompt passes all 10 checks: route to image/video generation tool immediately. no further review.

if prompt fails any single check: do not generate. instead, return the prompt to the author with a list of which steps failed and why (include the specific vague phrase or missing element). request revision and resubmit for re-audit. do not skip failed items.

if no character reference image is provided (step 2): mark step 2 as ❌ FAIL. a prompt cannot lock subject consistency without a reference asset or at minimum a detailed description + costume code. do not proceed.

if no location reference is provided (step 3): mark step 3 as conditional PASS only if the prompt names ≥3 concrete architectural or prop anchors. if location is abstract or single-sentence, mark ❌ FAIL.

if light direction is described as "natural" or "ambient" (step 4): mark ❌ FAIL regardless of other wording. these terms are undefined. require explicit direction rewrite.

if shadow direction cannot be determined from light direction (step 9): mark ❌ FAIL. the two must be explicitly linked. do not infer.

if the prompt is for a batch of shots and cross-shot consistency is a concern: after passing all 10 checks per prompt, run a secondary check: compare character appearance, costume, location continuity, and light continuity across consecutive shots in the batch. flag any breaks (e.g., character skin tone shifts, lighting reversal between adjacent shots). log as BATCH CONSISTENCY FAIL and require revision.

if generation tool or asset management system is down or unreachable: the checklist itself is synchronous and offline. proceed with audit; log the asset paths and regenerate later when systems are available.

output contract

on pass:

• return a marked checklist grid (10 items, all ✅)
• append the prompt to the output with a green "READY FOR GENERATION" label
• store the checked prompt and its [ASSET] block in a retrievable log (spreadsheet, database, or text file) for reproducibility and batch consistency reviews

on fail:

• return a marked checklist grid with ❌ next to each failed item
• include a bulleted "REQUIRED REVISIONS" section listing the exact item number, current text, and what is missing or incorrect
• do not append the prompt to any generation queue
• flag the prompt status as BLOCKED or PENDING REVISION in any tracking system

format examples:

PASS OUTPUT:
✅ ① Day/Night: "night" explicitly stated in opening clause
✅ ② Character: char_ref_001.png, white qipao, bun, earring detail
✅ ③ Space: mahogany desk (center), banker lamp (left), blinds (right wall)
✅ ④ Light: left 45°, hard rim, no fill
✅ ⑤ Ground Shadow: beneath feet, stretching 1.5m right
✅ ⑥ Texture: realistic pores, matte/semi-matte unified
✅ ⑦ Depth: blurred bookshelf foreground, character midground, window background
✅ ⑧ Shot Type: medium shot aligns with narrative focus
✅ ⑨ Projection: shadow right opposes light left 45°
✅ ⑩ Assets: [ASSET] char: char_001.png | location: office_int_003.jpg | prop: lamp_002.png

STATUS: READY FOR GENERATION
PROMPT: [original prompt text]

---

FAIL OUTPUT:
✅ ① Day/Night: "night" explicit
❌ ② Character: vague ("beautiful woman"), no reference asset, no costume code
✅ ③ Space: desk, lamp, window named
❌ ④ Light: says "soft natural lighting" (undefined), no angle given
✅ ⑤ Ground Shadow: present
✅ ⑥ Texture: unified
✅ ⑦ Depth: three layers present
✅ ⑧ Shot Type: medium shot appropriate
❌ ⑨ Projection: light direction vague, cannot verify shadow opposition
❌ ⑩ Assets: no [ASSET] block

STATUS: BLOCKED - REVISIONS REQUIRED
REQUIRED REVISIONS:
- Step 2: replace "beautiful woman" with specific description (age, hair color, hairstyle, facial expression, body pose, clothing fabric and fit). provide character reference image or asset code.
- Step 4: replace "soft natural lighting" with concrete direction, e.g., "single light from left at 45 degrees" or "overhead fill with left-side key light."
- Step 9: once light direction is fixed, add explicit shadow description matching the angle (e.g., "shadow stretching to right").
- Step 10: add [ASSET] block at end with character, location, and prop file paths.

RESUBMIT AFTER REVISION FOR RE-AUDIT.

outcome signal

user will know the skill worked when:

1. a prompt marked PASS generates images/videos with consistent lighting direction across shots in a batch.
2. a prompt marked PASS shows no floating characters, clear ground contact, and spatial depth.
3. a prompt marked PASS generates character appearances matching the reference asset in step 2 across multiple regenerations.
4. a prompt marked FAIL is returned to the author with specific line-item feedback, and the author can revise and resubmit without guessing what went wrong.
5. after running the checklist on a batch of 10 shots, cross-shot consistency issues (e.g., character skin tone, costume, lighting direction flip between adjacent shots) are caught before generation and flagged for revision.
6. weekly error logs show a declining rate of FAIL marks on previously-common steps (e.g., vague lighting language in step 4), indicating the team is internalizing the rules.
7. generation output shows improved spatial realism: clear shadows, depth layers, motivated shot types, and consistent character/environment continuity across shots.

---

quick reference card

print or post this checklist at your workstation:

✅ ① TIME: first clause states day / night / dawn / dusk / golden hour
✅ ② CHARACTER: matches reference image, name + ≥3 concrete details + asset path
✅ ③ SPACE: ≥3 named hard objects (desk, lamp, window, wall color, floor type)
✅ ④ LIGHT: explicit direction (left / right / 45° / overhead), zero "natural/soft/ambient"
✅ ⑤ SHADOW: ground contact shadow described, direction stated
✅ ⑥ TEXTURE: no retouching words (smooth, porcelain, flawless), materials unified
✅ ⑦ DEPTH: foreground element + character midground + background environment
✅ ⑧ SHOT TYPE: wide/medium/close matches emotional intent
✅ ⑨ PROJECTION: shadow direction opposes light angle, physics consistent
✅ ⑩ ASSET: [ASSET] block at end, all paths standardized

RESULT: 10/10 = PASS (generate) | <10/10 = FAIL (revise and resubmit)

---

credits: original prompt audit framework by beermanzz (clawhub). adapted and enriched for implexa standards with explicit decision points, asset management guidance, and cross-shot consistency protocols.