midjourney-prompt-auditor

Midjourney Prompt Auditor

Evaluate Midjourney prompts against seven structural criteria before generation. Produce a scorecard with pass/fail verdicts, specific failure explanations, and surgical rewrites for anything that fails.

When to Use

• User submits a prompt and wants it reviewed before generating
• User got bad results and wants to understand why
• User asks to "optimize", "fix", "check", "audit", or "improve" a prompt
• User asks "will this work?" or "what's wrong with this?"
• As a QA pass after generating prompts with midjourney-prompt-generator or midjourney-sensual-art
• When batch-preparing prompts and wanting consistency checks

The Seven Audit Criteria

Run every prompt through all seven checks, in order. Each check produces PASS or FAIL plus a specific explanation.

1. Token Competition

What it catches: Too many descriptors fighting for attention, causing Midjourney to pick winners semi-randomly.

How to check:

• Identify the three concepts Midjourney will weight most heavily (first ~10 words, plus any high-CLIP-signal terms anywhere in prompt)
• Ask: are those three the user's actual priorities?
• Count total descriptive concepts (not articles/prepositions). More than 12-15 distinct concepts = competition risk
• Check if any single concept is described with redundant synonyms ("messy, tousled, disheveled, unkempt hair" — pick one)

PASS if: The top 3 weighted concepts match user intent, total concept count is manageable, no redundant synonyms FAIL if: Important concepts are buried past word 30, or the prompt has 15+ competing concepts, or the same idea is restated in different words

Rewrite strategy: Front-load the priority concepts. Merge or cut redundant descriptors. Move secondary details after primary subject+style.

2. Grammar Conflicts

What it catches: The prompt asking for two incompatible rendering paradigms simultaneously.

Common conflict pairs:

• Camera language ("85mm", "f/2.8", "bokeh", "lens flare") vs. painting language ("impasto", "brushstrokes", "oil on canvas")
• "Clean white backdrop" vs. "chiaroscuro" or "dramatic shadows"
• "Photorealistic" vs. "watercolor" or "sketch"
• "Minimalist" vs. detailed scene descriptions
• "Flat design" vs. volumetric lighting
• "Candid snapshot" vs. elaborate posed composition
• Modern clothing/setting terms vs. historical/fantasy period terms
• Multiple art movements that pull in opposite directions ("art nouveau AND brutalist")

Productive tension (NOT a conflict):

• "Photorealistic face dissolving into gestural abstraction" — achievable because it defines WHERE each paradigm applies
• "Digital painting with visible oil brushwork" — these coexist because one is medium, one is texture

The test: For each pair of descriptors, do they pull the model toward the same rendering pipeline or different ones? If different — does the prompt specify HOW they coexist spatially, or is Midjourney left to guess?

PASS if: All descriptors are compatible, or tensions are explicitly resolved with spatial/hierarchical instructions FAIL if: Incompatible paradigms coexist without resolution

Rewrite strategy: Pick one paradigm. If the tension is intentional, add explicit spatial instructions ("face rendered with photographic precision, background dissolving into loose brushwork").

3. CLIP Vocabulary Alignment

What it catches: Terms the model doesn't know, poetic phrases with no visual correlate, human-readable mood words that CLIP maps to nothing or something random.

High-CLIP-signal terms (model knows these well):

• Established art movements: "impressionist", "art nouveau", "baroque"
• Named techniques: "chiaroscuro", "impasto", "sfumato", "tenebrism"
• Known artists with large training representation: "Sargent", "Caravaggio", "Mucha", "Frazetta"
• Photography terms: "rim light", "golden hour", "shallow depth of field"
• Material descriptors: "oil on canvas", "watercolor", "charcoal", "ink wash"
• Standard composition terms: "rule of thirds", "dutch angle", "bird's eye view"

Low-CLIP-signal terms (likely ignored or randomly mapped):

• Abstract emotional phrases: "the quiet ache of longing", "sovereign stillness"
• Invented compound descriptors: "audace silencieuse", "naughty innocence"
• Narrative/temporal concepts: "the aftermath more eloquent than the act"
• Non-visual sensory language: "the taste of copper", "a whispered secret"
• Obscure or niche references the model likely hasn't seen in captions
• French or non-English terms (unless they're art-specific like "trompe l'oeil")

Gray zone — might work, test first:

• Mood words with some visual correlate: "melancholic", "ethereal", "ominous"
• Cultural references: "Victorian", "Edo period", "Moorish"
• Emotional states with visual expression: "defiant", "vulnerable", "smug"
• Specific-but-niche artists: test with --s 0 to see if recognized

PASS if: 80%+ of descriptive terms have strong CLIP signal. Low-signal terms are limited to 1-2 closing flourishes (harmless dead weight) FAIL if: Core structural terms (subject, style, lighting) rely on low-signal vocabulary, or more than 30% of the prompt is CLIP-invisible

Rewrite strategy: Replace low-signal terms with high-signal equivalents. Keep poetic closers only if they don't compete with functional terms. Example: "sovereign stillness" → "composed regal expression, steady gaze" (two CLIP-visible terms instead of one invisible one).

4. Negative Prompt Coherence

What it catches: The --no flag contradicting the positive prompt, or missing critical exclusions.

How to check:

• List every term in --no
• For each, scan the positive prompt for the same term or a strong synonym/implication
• Check: does the positive prompt use camera language while --no blocks "photograph"? (The positive terms likely overpower)
• Check: are there obvious exclusions missing? (e.g., a painterly prompt with no --no text watermark)

Common contradictions:

• Prompt says "warm amber light" + --no warm
• Prompt uses "85mm shallow depth" + --no photograph
• Prompt describes "detailed background" + --no background

Common missing exclusions:

• Painterly prompts should typically --no text watermark
• Character prompts often need --no extra limbs duplicate
• Dark/moody prompts may need --no bright neon oversaturated

PASS if: No contradictions between positive and negative, reasonable exclusions present FAIL if: Any term appears in both positive and --no, or critical exclusions are missing for the style

Rewrite strategy: Remove contradictions (decide which direction wins). Add standard exclusions for the genre.

5. Specificity Gradient

What it catches: Uniform detail density that fights the model, or critical elements left unspecified while trivial ones are over-described.

Principle: Detail should cluster on what the user cares about controlling. Deliberately vague areas give Midjourney room to improvise (which it's good at).

How to check:

• Identify the three most detailed sections of the prompt
• Ask: are those the three things the user would be most upset to see wrong?
• Check for "orphan specifics" — one hyper-detailed element surrounded by vague ones (e.g., "a woman in a room with a specifically described 14th century Venetian glass chandelier with 47 candles" — the chandelier will dominate the image)
• Check for "missing anchors" — no description at all for a critical element (e.g., detailed face description but zero guidance on what she's wearing → random modern clothing on a fantasy character)

PASS if: Detail density correlates with importance. Critical elements are specified, secondary elements are loosely guided or omitted FAIL if: Detail is uniformly distributed, or the wrong things are specified, or a critical element has zero guidance

Rewrite strategy: Add minimal guidance to unspecified critical elements. Reduce over-specified secondary elements to 1-2 words. Let the specificity gradient match priority.

6. Parameter–Content Alignment

What it catches: Technical parameters that fight the prompt's intent.

Key checks:

Parameter	Check
--ar	Does it match the composition? Portrait prompt + --ar 16:9 forces unwanted horizontal fill. Full-body character needs --ar 2:3, not --ar 1:1.
--s	Stylize 0-50 fights painterly/artistic prompts. Stylize 500+ fights prompts needing precision/accuracy. Sweet spots: photorealistic 50-100, painterly 200-350, experimental 500+.
--c	Chaos >30 on prompts requiring consistency (character sheets, series). Chaos <10 on exploration prompts wastes the opportunity for variation.
--v	V7 for most work. V6 if using --cref, multi-prompting with weights, or needing legacy quality values.
--q	V7 uses 1/2/4. If prompt says --q 0.5 on v7, it's invalid.
--no	See criterion 4 above.
--raw	Fights high-stylize. Redundant with --s 0.

PASS if: All parameters support the prompt's intent and are valid for the selected version FAIL if: Any parameter contradicts intent or is invalid for the version

Rewrite strategy: Adjust parameters to match intent. Flag version-incompatible values.

7. Reduction Test

What it catches: Dead weight — words that take up tokens without influencing the output.

How to check:

• Read each phrase in the prompt
• Ask: if I remove this phrase, does the imagined output change?
• If no → it's dead weight
• Categories of dead weight:
  • Redundant restating: "same blue eyes, consistent blue eyes, matching blue eyes" → say it once
  • Human-readable-only closers: "the painting itself an act of slow attention" → CLIP doesn't parse this
  • Filler adjectives: "very", "really", "truly", "extremely" → CLIP ignores these
  • Narrative framing: "imagine a scene where", "picture this:" → invisible to model
  • Contradicted terms: anything also listed in --no

PASS if: Removing 20% or less of the prompt would change nothing. Prompt is lean. FAIL if: 30%+ of the prompt is dead weight. There's room for significant compression.

Rewrite strategy: Cut dead weight. Compress redundancies. A lean 50-word prompt where every token pulls its weight outperforms a verbose 120-word prompt.

Output Format

For each audit, produce this scorecard:

## Prompt Audit Scorecard

**Prompt:** [the prompt being audited, truncated to first 50 words + "..."]
**Version detected:** [v5/v6/v7 or "not specified"]
**Word count:** [N words positive prompt] + [N words in --no]

| # | Criterion | Verdict | Issue |
|---|-----------|---------|-------|
| 1 | Token Competition | PASS/FAIL | [specific issue or "—"] |
| 2 | Grammar Conflicts | PASS/FAIL | [specific issue or "—"] |
| 3 | CLIP Vocabulary | PASS/FAIL | [specific issue or "—"] |
| 4 | Negative Coherence | PASS/FAIL | [specific issue or "—"] |
| 5 | Specificity Gradient | PASS/FAIL | [specific issue or "—"] |
| 6 | Parameter Alignment | PASS/FAIL | [specific issue or "—"] |
| 7 | Reduction Test | PASS/FAIL | [N% removable] |

**Score: [N]/7**

### Failures Explained
[For each FAIL: what specifically is wrong and why it matters for the output]

### Rewritten Prompt
[Full rewritten prompt with all failures addressed. Show the prompt ready to paste.]

### What Changed
[Bullet list of specific changes and rationale — not vague "improved flow" but "moved X before Y because weight distribution", "replaced Z with W because CLIP signal"]

Style-Specific Audit Extensions

When the user is working in a known house style (e.g., Fantasy Vixens painterly impasto), apply additional checks:

Fantasy Vixens House Style Checklist:

• Does the prompt include impasto/palette knife texture language?
• Is chiaroscuro lighting specified with warm single source?
• Is the palette limited to 3 hues maximum?
• Are edges described with lost-and-found hierarchy (hard on focal points, soft elsewhere)?
• Does --no exclude "photograph digital smooth clean airbrushed flat background text watermark"?
• Is --s in the 250-300 range?
• Is --ar 2:3 for portraits?
• Is the face/skin rendered with more precision than the background (specificity gradient matching the style's core paradox)?

If working in this style, add a House Style Compliance row to the scorecard (PASS/FAIL) with specific missing elements listed.

Batch Audit Mode

When auditing multiple prompts (e.g., a series or character sheet set):

1. Run the 7-check audit on each prompt individually
2. Then run cross-prompt checks:
   • Palette consistency: Are the same character's colors consistent across prompts?
   • Parameter consistency: Are --s, --c, --v the same across the batch?
   • Identity anchors: Do character-defining features (eye color, hair, distinguishing marks) appear in the same position and wording across all prompts?
   • Style coherence: Are all prompts using the same rendering paradigm?
3. Flag any cross-prompt inconsistencies in a separate section of the report

Edge Cases

• Intentionally experimental prompts: If --c is high (50+) and the prompt is clearly exploratory, relax criteria 1, 5, and 7. The user is exploring, not controlling.
• Multi-prompt syntax (:: weights): Check that weights are valid for the version. V7 has limited multi-prompt support — flag if v7 is specified.
• Style reference prompts (--sref): When a style reference URL is present, the prompt's style language matters less (the reference dominates). Relax criterion 3 for style terms but keep it strict for subject terms.
• Permutation prompts ({option1, option2}): Audit each permutation independently if there are ≤4. If more, audit the base structure and spot-check 2-3 permutations.