6. Getting quality results through iteration

When Abebe, a graphic designer in Addis Ababa, first tried an AI image generator, he typed "logo for a coffee shop" and was disappointed by the generic result. He assumed AI was not useful for professional work. Three months later, after learning to iterate, he uses the same tool to produce client-ready assets in under 30 minutes.

The difference was not the tool — it was the process.

Generative AI works like a very talented but very literal assistant. It will do exactly what you ask, nothing more and nothing less. The challenge is that most people ask in a way that is too vague, too short, or misses critical details. The AI fills those gaps with its own guesses — which are often wrong for your specific context.

Iteration means treating AI output as a draft, not a final product. Professional designers, copywriters, and video producers who use AI effectively all follow the same pattern:

1. Generate a first version quickly

2. Evaluate it against specific criteria

3. Identify exactly what is missing or wrong

4. Refine the prompt (or edit the output directly)

5. Generate again

6. Repeat until the result meets your standard

This loop — generate, evaluate, refine — is the core skill. Once you internalise it, your results improve dramatically and consistently.

How do you know what to change when a result is not good enough? A structured evaluation framework helps. Use the CRAFT checklist after every AI output:

C — CLARITY: Is the result clearly communicating the right message? Does a viewer immediately understand what the image, text, or audio is about?

R — RELEVANCE: Is the output relevant to your specific audience and context? A generic marketing image might be clear but completely wrong for Ethiopian customers who do not recognise the people, food, or setting in it.

A — ACCURACY: Are all facts, prices, names, product details, and dates correct? AI sometimes "hallucinate" incorrect information confidently — you must verify every factual claim.

F — FORMAT: Is the output in the right format? Is an image the right size and aspect ratio? Is the text the right length? Does the audio hit the right duration? Is the video in landscape or portrait as needed?

T — TONE: Does the output match your brand tone? A children's education app needs a warm, playful tone. A financial services company needs a serious, trustworthy tone. Mismatched tone is one of the most common problems in AI-generated content.

After scoring each dimension (Good / Needs work / Wrong), you know exactly which part of your prompt to improve. If RELEVANCE fails, add local context. If TONE is wrong, specify the emotion explicitly. If FORMAT is off, add dimension or length constraints.

Let us walk through a real iteration example. Sara runs a small textile shop in Merkato, Addis Ababa, and wants marketing copy for an Instagram post promoting a new collection of habesha kemis (traditional Ethiopian dress).

── ATTEMPT 1 (first prompt) ──

Prompt: "Write Instagram caption for habesha kemis sale."

Result: "Step into tradition! Our Habesha Kemis collection is here. Shop now and celebrate your culture. Limited stock available!"

CRAFT evaluation:

• Clarity: OK — the topic is clear.

• Relevance: Weak — no mention of Merkato, no price hint, no Ethiopian tone.

• Accuracy: No facts to check yet.

• Format: OK — correct length for Instagram.

• Tone: Generic, not warm or personal.

── ATTEMPT 2 (refined prompt) ──

Prompt: "Write an Instagram caption in a warm, personal Ethiopian tone for a habesha kemis sale at Sara's Textile Shop in Merkato, Addis Ababa. Prices start from 1,500 birr. The collection uses hand-woven Awash cotton. Include a TeleBirr payment option and a call to action to DM the shop. Max 3 sentences, end with relevant emojis."

Result: "አዲሱ የሐበሻ ቀሚስ ስብስባቻችን ደረሰ! ከ1,500 ብር ጀምሮ — ከአዋሽ ጥጥ ዘንድ ፈትሎ፣ ሐቀኛ ጥበብ ዘንድ ፈትሎ። ቴሌብርር ወይም በቀጥታ ሱቃቻን DM አድርጉን — ለሚቀጥሉ 48 ሰዓቶቹ ነጻ ዴሊቨሪ! ✨👗🇪🇹"

CRAFT evaluation:

• Clarity: Excellent — product, price, and action are clear.

• Relevance: Strong — Amharic, birr, local shop, TeleBirr.

• Accuracy: Sara should verify the 48-hour delivery promise matches her actual offer.

• Format: Good — three sentences, emojis at the end.

• Tone: Warm, personal, authentic.

The jump from Attempt 1 to Attempt 2 required just two minutes of extra thought. The quality improvement was enormous. This is what systematic iteration looks like.

One risk with iteration is over-iteration: spending so long refining a prompt that you lose more time than the AI ever saved you. Professional AI users learn to recognise the "good enough" threshold.

Here are the signs that you should stop iterating and ship the output:

1. THE OUTPUT PASSES ALL FIVE CRAFT DIMENSIONS at a level appropriate for its purpose. A social media post does not need to be as polished as a billboard ad.

2. FURTHER CHANGES PRODUCE NO NET IMPROVEMENT. If your last two attempts produced results that are different but not clearly better, you have hit the ceiling of what prompting alone can achieve.

3. IT WOULD BE FASTER TO EDIT DIRECTLY. If the AI output is 90% correct and the remaining 10% is a small text tweak, a colour change, or a crop — just fix it in the editing tool. Do not keep re-prompting for a minor change.

4. DEADLINE PRESSURE. Iteration is a tool for quality, but professional content creation also has schedules. A good result delivered on time beats a perfect result delivered late.

A practical rule: for most marketing content, allow yourself 3 to 5 prompt iterations maximum. If you are still not satisfied after 5 attempts, step back and ask whether the underlying brief (what you are trying to say and to whom) is clear enough. Often the problem is not the AI — it is an unclear brief.

Remember: AI is your drafting tool, not your quality control department. Human review and judgement are always the final step.

Iteration looks slightly different depending on whether you are working with images, text, or audio. Here are the most important tool-specific tactics.

ITERATING ON AI IMAGES (Midjourney, DALL-E, Adobe Firefly):

• Use the "vary" or "remix" feature instead of rewriting the full prompt — change only the part that needs fixing.

• Add negative prompts to exclude unwanted elements. Example: if the AI keeps generating Western-looking food on the plate, add "—no pizza, no burger, no bread" and specify "injera, tibs, shiro".

• Lock the elements you like with a seed number (most tools show the seed in the output panel) so that repeated generations maintain the same base composition.

• Change one variable at a time. If you change the lighting, style, and subject together, you will not know which change improved (or worsened) the result.

ITERATING ON AI COPY (ChatGPT, Claude, Gemini):

• Use follow-up instructions rather than rewriting the full prompt. After receiving a result, type: "Keep the same structure but make the tone 30% warmer and add a TeleBirr payment mention in the last sentence."

• Ask for multiple options. End your prompt with "give me 3 versions" — then pick the best elements from each and ask for a final synthesis.

• Use "act as" to align tone: "Act as a friendly Addis Ababa market vendor writing to regular customers."

ITERATING ON AI AUDIO (ElevenLabs, Suno):

• For voice-overs: if the pace is wrong, use the speed slider rather than rewriting the script.

• For music: change one element per generation — first get the mood right, then the tempo, then the instrumentation. Changing everything at once wastes iterations.

• Save every generated version before running the next — you may prefer version 2 over version 4 when you listen again with fresh ears.