ai-drawing-prompt

AI Drawing Prompt

Build prompts as a reproducible package, not a single line. Load only the references needed for the target model.

1. Lock The Prompt Contract

Capture and confirm:

1. Subject and scene.
2. Visual goal (mood, style, era, lighting, composition).
3. Delivery target (model, aspect ratio, quality/speed preference, output format).
4. Hard exclusions (objects, artifacts, unsafe elements, brand/legal constraints).
5. Number of variants and iteration budget.
6. Reference image plan (optional): source image(s), what to preserve, what must change, and strength (low/medium/high).

If the model is unknown, propose one model choice plus one fallback.

2. Load Model References

Always read:

• references/model_profiles.md
• references/safety_and_negative_guidance.md

Read when needed:

• references/iteration_playbook.md for troubleshooting and iterative refinement.
• references/source_selection.md when provider rules may have changed and source confidence matters.

3. Build The Prompt Package

Produce all blocks:

1. Primary prompt: final production prompt.
2. Parameter block: model-specific controls only.
3. Negative/Avoidance block: provider-correct form (or explicit "not supported").
4. Reference image block (if provided): adapter instructions and preserve/change list.
5. Variant set: at least 3 controlled variants (style, composition, lighting, camera distance, or detail level).
6. Iteration plan: 2-3 rounds with expected failure checks and next edits.

4. Enforce Model-Correct Constraints

Apply these hard rules:

• Do not invent unsupported parameters.
• Keep parameter names and ranges aligned with the selected provider profile.
• FLUX: do not emit classic negative prompts as a separate field.
• Midjourney: use --no only for allowed, concrete exclusions.
• SDXL: use prompt weights for de-emphasis only when the caller supports weighted prompts.
• OpenAI GPT Image: encode avoidances in natural-language prompt constraints.
• When reference image is provided, explicitly separate preserve from change instructions.

5. Return In A Standard Output Shape

Use this structure in responses:

### Prompt Pack
Model: <provider/model>
Intent summary: <1-2 lines>

Primary prompt:
<text>

Parameter block:
<provider-specific key/value or CLI parameters>

Negative/Avoidance block:
<model-correct form or "Not supported by model; constraints folded into main prompt.">

Reference image block:
<optional: input image usage, preserve list, change list, strength>

Variants:
1) <variant A>
2) <variant B>
3) <variant C>

Iteration plan:
Round 1 -> <what to evaluate>
Round 2 -> <what to adjust>
Round 3 -> <final polish controls>

6. Safety Boundary

If the request is unsafe, illegal, or policy-violating:

1. Refuse the harmful part directly.
2. Offer a safe alternative prompt objective.
3. Continue with compliant prompt construction.