Optimize Then Generate — Plan, Evaluate, Improve Pipeline

Why "Optimize First" Matters

Most failed AI generations aren't model problems—they're prompt problems. A rough sentence like "nice product photo" leaves lighting, angle, background, and style up to chance. AI Chat turns vague intent into structured instructions that image and TTS models can execute reliably.

For overseas ecommerce and performance teams, the cost of skipping this step is concrete: burned credits on 4K re-rolls, Amazon mains that fail compliance checks, and Meta/TikTok ads that look “AI-random” instead of on-brand.

After optimization you typically get:

• Clearer intent — subject, scene, channel, and goal are explicit
• Style consistency — same brand look across dozens of SKUs or ad variants
• Detail control — texture, lighting, and composition are named
• Predictable outputs — fewer random failures and re-rolls
• Credit efficiency — filter at 1K, finalize at 2K/4K only for winners

Before and After Example

Rough input (unstable):

Nice skincare bottle photo for ads

After AI Chat optimization (stable):

Single skincare serum bottle, centered hero product, pure white seamless background,
soft studio key light from upper left, subtle contact shadow beneath bottle,
glass texture with realistic reflections, label text sharp and readable,
premium DTC aesthetic, photorealistic, 4:5 aspect ratio for Instagram feed ad

The optimized version names subject placement, background, lighting direction, material, and constraints—so Nano Banana 2 and related models have less room to guess wrong.

Plan–Evaluate–Improve loop

Compress the five steps into three memorable actions:

Teams often skip Evaluate and re-roll ten times. A short checklist (subject readable, color match, artifacts, voice pacing) separates prompt structure issues from model/resolution choices—saving credits on blind retries.

The same loop applies to conversational editing: reference upload → first pass "lighting only" → evaluate → second pass "subtle skin"—never stack every retouch term in one prompt.

The Core Pipeline (5 Steps)

1. Draft a rough prompt

Write what you want in plain language. Don't worry about structure yet.

2. Refine in AI Chat

Paste the rough idea and ask for structured variants. Request 3 style directions—for example: minimal studio, lifestyle natural light, and bold campaign color.

Compare variants for:

• Visual clarity (is the product the hero?)
• Brand fit (colors, mood, premium vs playful)
• Keyword conflicts (avoid mixing "photorealistic" and "flat illustration" in one prompt)

For voice scripts, ask AI Chat to shorten sentences and add a clear hook + CTA.

Each chat reply costs 0.5 credits and saves to Prompt Library.

3. Generate image or voice

Pick one variant and send it to AI Image or paste a polished script into AI Voice.

AI Image credits by resolution: 1K = 3, 2K = 4, 4K = 8 credits per generation.

AI Voice: 1 credit per 500 characters (minimum 500).

For voice, generate a short sample first (10–20 seconds) to validate tone and pacing before the full read.

4. Compare and iterate

Score outputs on a simple rubric:

Adjust one variable at a time—lighting, background, voice choice, or script length—not everything at once.

5. Save as reusable template

Store the winning prompt or script with metadata:

• Use case (listing, ad, social cover, voiceover)
• Aspect ratio (1:1, 4:5, 9:16) for images
• Model notes (nano-banana-2 for quality, nano-banana-fast for drafts)

Next time you only swap the product name or scene detail.

Model Selection Quick Reference

When to Chat First vs Refine Existing Prompt

See AI Chat Guide for optimization request templates.

Production SOP for agencies and in-house teams

Use this when more than one person touches creatives (freelancer + brand, or a small growth pod):

1. Brief in one sentence — channel (Amazon main / Meta 4:5 / TikTok 9:16), deliverable count, hard constraints (white BG, no props, preserve label).
2. Chat owner drafts 3 structured variants; mark one as primary.
3. Image owner runs 1K screen (4–6 variants) → scores with the rubric above → promotes 1–2 to 2K.
4. Voice owner (if needed) samples 10–20s before full TTS.
5. Librarian tags the winner in Prompt Library: channel + SKU family + model + ratio.

Skipping step 5 is how brand look drifts by week three. The Prompt Library is the system of record—not a Slack screenshot of a Discord prompt.

Credit math: why chat-first usually wins

Assume a new Meta 4:5 ad brief with no library template yet:

The chat-first path is rarely more expensive and usually cheaper once you count the re-rolls you didn’t need. For SKU batches with a gold template, skip exploration and only swap fields (see Ecommerce Image Optimization).

Team Workflow: Shared Prompt Library

For ecommerce, UGC ads, or social teams, one shared library beats everyone prompting from scratch:

Review templates monthly. Retire prompts that consistently underperform in CTR or conversion tests.

End-to-end case: brief to ad-ready assets (90 min)

Scenario: Wireless earbuds TikTok 9:16 ad—cover still + 20-second voice read.

Timeline
09:00  AI Chat — brief + 3 visual directions + 2 hook scripts        (~15 min, 2 credits)
09:15  AI Image — fast 1K × 6 cover filter                             (~20 min, 18 credits)
09:35  Checklist score — pick "UGC handheld + window light"             (~5 min)
09:40  AI Image — nano-banana-2 2K finals × 2                          (~10 min, 8 credits)
09:50  AI Chat — 40-word script, spoken phrasing                        (~5 min, 0.5 credits)
09:55  AI Voice — 15s samples × 2 voices                               (~10 min, 2 credits)
10:05  Editor — still + VO + captions                                  (~25 min)

Evaluate log (fill every run):

Failed rows map to one lever only—do not change prompt, model, and aspect ratio at once.

Common Mistakes

• Skipping chat refinement on "simple" product shots—background and lighting still vary wildly
• Changing too many keywords between iterations—you won't know what fixed the output
• Ignoring aspect ratio until export—compose for 9:16 or 1:1 from the prompt stage
• Long voice scripts on first try—validate tone on a short clip before the full read

FAQ

Does optimization work for AI Voice too?
Yes. The same structured hook + benefit + CTA pattern applies to ad reads, explainers, and social voiceovers.

How many variants should I generate?
Three optimized variants plus 1–2 manual tweaks is enough for most decisions. More than five slows you down without better results.

Can I reuse one prompt across models?
Use the same structure; adjust resolution and model-specific quality terms as needed.

Related Guides

• AI Chat Guide
• Ecommerce Product Image Optimization
• AI Voice Workflow
• Social Media Batch Creative