| name | mu-prompt-nudger |
| description | Revise AI image/video prompts with precise μ nudges that escape centered model defaults while preserving fidelity. Use when a prompt or generated result is too centered, too frontal, too filled, too blank, overdrifted, over-stylized, or losing identity, and the user needs a one-step correction using subject offset, luminance bias, gaze angle, footprint, negative-space ratio, symmetry caps, and fidelity locks. |
Mu Prompt Nudger
Purpose
Use this skill to rewrite prompts so they move toward μ: off-center enough to break the model's safe statistical default, but controlled enough to preserve identity, realism, lighting, and coherence.
This skill is independent. It includes its own nudge bands, templates, and helper script. It does not require mu-finder or any other skill.
Quick Start
Generate a correction clause from symptoms:
python3 scripts/mu_nudge.py --symptom centered --symptom frontal --symptom crowded
Generate a correction from measurements:
python3 scripts/mu_nudge.py --db 3 --dg 1 --sr 55 --nr 25 --fidelity 8
Use references/nudge-playbook.md for detailed symptom-to-correction mappings and prompt templates.
Core Nudge Bands
- Subject offset
ΔB: target 10-12% from geometric center.
- Luminance offset
ΔL: target 6-9% toward the long side.
- Gaze/dominant direction
ΔG: target 6-8 degrees toward the long side.
- Subject footprint
S_r: target about 40% of frame width.
- Negative space
N_r: target 40-50% of frame area.
- Fidelity lock: preserve base identity, lighting direction, texture, and color temperature within about 10%.
Effectiveness
After applying a nudge: re-measure (μ Finder or MOS)
compare: Did ΔB / ΔL / ΔG actually move?
If not: Nudge Ignored → increase constraint strength
Conflict Detection
Example conflicts:
- “expand void” + “increase subject footprint”
- “increase offset” + “restore fidelity” (too aggressively)
If conflict: Conflicting Nudges → prioritize primary, delay secondary
Primary Lock
Only one axis should dominate:
- Primary: ΔB (offset)
- Secondary: ΔG (gaze)
If multiple axes shift equally: Diffuse Correction → weak result
Strength Levels
- soft → near band
- standard → default
- hard → resistant models
Example:
- soft: 8–10%
- standard: 10–12%
- hard: 12–15%
Resistance
If:
- repeated attempts → no movement
Model Resistance → switch to structured block or hard constraints
Translation Strength
Check if prompt includes:
- numeric control (good)
- spatial instruction (required)
- avoidance clause (critical)
If missing: Weak Prompt → likely ignored
Outcome Split
After generation: numeric: ΔB, ΔL, ΔG
visual: does it feel off-center?
If mismatch: Numeric Success / Visual Failure → refine constraints
Compensation
Detect:
- offset present
but:
- lighting recenters
- props rebalance
- framing compensates
Compensation Detected → increase multi-axis enforcement
Void Type
Check if void is:
- active (good)
- empty (bad)
- cluttered (bad)
If wrong type: Void Misclassification → wrong nudge applied
Iteration Rule
- 1 nudge → generate → measure → adjust
If violated: Stacked Nudges → attribution loss
Repair Mode
Given an image:
- identify:
- output:
- Remove X
- Increase Y
- Preserve Z
Not just rewrite prompt—surgical correction
Proof Validation
After applying nudge:
- run Proof Loop
If not proven:
- Unproven Nudge → aesthetic bias
Workflow
- Identify the failure. Classify the current prompt/result as centered, frontal, crowded, blank, overdrifted, over-stylized, prop-heavy, symmetry-safe, or fidelity-broken.
- Choose the smallest correction but only if the model actually responds to it. Apply one primary μ nudge first. Add at most two supporting locks.
- Preserve the subject class. Do not rewrite the entire prompt unless the original is unusable.
- Add numeric control. Use explicit offsets, ratios, and caps when the model keeps snapping back to center.
- Add fidelity locks. Preserve base tonality, identity, lighting direction, texture, and realism before increasing drift.
- Return a revised prompt. Keep the original subject and style, but append or integrate a μ-control clause.
Output Format
Return:
- Diagnosis: what is failing and why.
- Primary Nudge: the one most important correction.
- Revised Prompt: a clean prompt ready to paste.
- Optional Explicit Controls: CAD-style values if the target model accepts structured prompts.
- Avoid: negations that prevent collapse, such as no centered framing, no perfect symmetry, no direct front-facing gaze, no heavy props.
Default One-Line Nudge
Use this when the prompt is generally too safe or centered:
Adjust toward μ: shift the subject center 10-12% off geometric center toward the long side, turn gaze or dominant direction 6-8 degrees into the open space, keep subject footprint near 40% of frame width, expand negative space to 40-50%, preserve base lighting, texture, color, and identity within 10%, no centered framing or perfect symmetry.
Guardrails
- Do not add more story objects to solve a spatial problem. Add spatial pressure first.
- Do not use maximum drift as the answer. μ is a mid-shift, not a collapse.
- If fidelity is already weak, restore identity/lighting/texture before increasing offsets.
- If the void is blank, add low-contrast gradient, edge trace, or peripheral anchor rather than filling it with props.
- If a prompt is already in μ, preserve it and make only small variations.
This package contains a modular visual reasoning skill suite built from Russell Parrish / A.rtist I.nfluencer protocols. The skills are designed to run independently, but they also interoperate through routing, handoff notes, and shared visual reasoning concepts. More information: www.artistinfluencer.com. Copyright 2026.