Iteration discipline: prompting as direction

The difference between someone who produces great AI-assisted creative work and someone who produces average stuff isn't the tool. It's the iteration discipline.

The amateur pattern

• Generate image.
• It's okay.
• Ship it.

Or:

• Generate image.
• Tweak prompt slightly.
• Generate again.
• Still not great.
• Give up.

Both produce mediocre output.

The craft pattern

• Develop a clear intent before prompting.
• Generate many variants (10-30).
• Study the variants; identify what works.
• Refine the prompt based on what the tool responded to.
• Generate next batch with tighter constraints.
• Converge on a final via 3-5 iteration rounds.

Takes more time. Produces meaningfully better output.

The "intent document"

Before any generation, write down:

• What is this for?
• What's the audience?
• What aesthetic are we going for?
• What must it convey?
• What must it avoid?

Five minutes of writing. Sharpens thinking. Makes the rest faster.

Iteration rounds

Round 1: wide exploration.

• Cast a broad net.
• Multiple styles, compositions, moods.
• Goal: find surprising directions.

Round 2: narrowing.

• Pick 2-3 directions from round 1.
• Refine prompts.
• Converge on style.

Round 3: refinement.

• One direction, many variants.
• Tighten specific elements.
• Polish.

Round 4: selection and finish.

• Pick the winner.
• Post-process (human editing).
• Ship.

Each round takes 30-60 minutes. Total: 2-3 hours for a good piece. Compared to amateur "ship the first okay one" — night and day quality.

Knowing when to stop

Two signals:

• Diminishing returns. Each new generation is a minor variation; no qualitative improvement.
• Intent satisfied. The output meets what you wrote in the intent document.

If you're still iterating at round 7, either you had the wrong intent or you're in perfectionism territory. Ship.

The reference discipline

Professional AI creative work leans on reference images / moodboards heavily:

• Pull 5-15 reference images before generating.
• Use them as style references in tools that support it.
• Describe specifics from them in prompts.

AI works better when you show it what good looks like, not just describe it.

Human editing is non-optional

Raw AI output has tells — subtle weirdness, imperfect details, occasional glitches. Edit:

• Image: Photoshop for blemishes, composition tweaks, color grading.
• Video: an editor for color, cuts, audio sync.
• Audio: a DAW for polish.
• Text: a human read for voice and accuracy.

The editing step is usually 20-40% of total effort and does most of the work of making output "yours."

Taste compounds

The more you iterate, the better your eye gets. Track:

• Which prompts produced outputs you loved.
• Which style references worked.
• Which genres of generation you handle well vs. struggle with.

A personal craft notebook is a career accelerant.

The "which to pick" problem

Staring at 20 candidates and not knowing which is best is real. Tactics:

• Walk away for 20 minutes. Return with fresh eyes.
• Show to a trusted colleague. Ask "which ones work?"
• Apply the intent document. Which one best matches what you wrote?
• Rank top 5. If it's hard to pick among them, any are acceptable — just ship.

The meta-lesson

Tools are tools. Quality work comes from discipline, craft, and iteration. Most "AI is cheap slop" critique is really a critique of people not iterating. The people who ship great work are investing as much time as traditional production; they're just able to explore more before committing.