Lesson 4. Constraints and quality criteria
Constraints don't make the answer poorer — they bring it closer to the result you actually need. In this lesson you'll see which constraints are useful and which ones stifle the prompt.
Key takeaways
- control length and tone
- properly add restrictions to a prompt
- check results using quality criteria
- separate required elements from undesirable patterns
Topic breakdown
When users tell the model 'answer freely', they often get output that's too long, too generic, or uncontrollable. This happens because the model wasn't given boundaries to work within.
Constraints are the prompt's internal control system. Length, tone, required blocks, restrictions, and quality criteria make the result noticeably more useful.
This is especially important for ads, support responses, SEO descriptions, and any repeatable content where you need not just text, but a controllable quality standard.
What you'll learn
- control length and tone
- properly add restrictions to a prompt
- check results using quality criteria
- separate required elements from undesirable patterns
Lesson plan
Why constraints are needed
They don't slow down the model — they help it work within the right range: shorter, more precise, and closer to your task.
Required blocks and restrictions
A strong prompt separately describes what must be included and what absolutely should not appear in the result.
Internal review by criteria
If you ask the model not only to write text but also to check it against criteria, quality becomes more consistent.
When constraints become too many
If constraints contradict each other or don't affect the result, they only confuse the model and reduce output quality.
Weak vs strong prompt
Write a short ad copy for our service. Make it really good.
You are a copywriter. Write a Telegram ad for an online accounting service. Audience: small business owners. Text — up to 120 words, tone simple and confident. Required blocks: headline, 2 benefits, CTA. Restrictions: clickbait, jargon, unsubstantiated promises. At the end, check the response for clarity, brevity, and CTA strength.
The strong prompt sets boundaries, structure, and verification criteria, so the result is closer to a working standard rather than a random text.
Deep dive
Constraints and criteria are the 'quality control' of a prompt. Without role, context and task the model does not know what to do. And without constraints it writes an answer with the best intentions that does not meet your standards. Professional prompt engineering pays special attention to these two elements.
The most effective constraint types: 1) Length — 'no more than 200 words' or '2-3 sentences per item'. 2) Tone — 'formal', 'friendly', 'technical' or 'in simple language'. 3) Bans — 'no generic advice', 'no answer without numbers'. 4) Required elements — 'an example in every item', 'a CTA in every answer'.
Quality criteria are pre-defined standards for judging the AI answer. For example: 'The answer must be actionable — the reader can take the first step right away'. Or: 'Every recommendation must fit a small business in the Fergana region'. The more precise the criterion, the more useful the AI result.
A practical rule: every prompt should have at least 2-3 constraints. But too many constraints (10+) confuse the model. The optimal formula: 1 length constraint + 1 tone + 1-2 required elements + 1 ban. These 4-5 constraints give the optimal result in almost any task.
Test lesson 4 with the constraints and criteria lab
Pick the constraints you need, turn on review criteria, and immediately see how controlled your prompt becomes.
1. Pick a situation
2. Manage constraints
3. Pick review criteria
Assembled prompt
You are a copywriter. Service: an online accounting service in Tashkent. Audience: small business owners. Task: write a short ad text for Telegram. The text must be no longer than 120 words. The tone must be simple and confident. Required blocks: headline, 2 key benefits and a CTA. At the end, briefly check the answer against the selected criteria. Evaluation criteria: clarity, audience fit, cta strength.
Quick audit
Constraint balance
Constraints are in a useful range. The result stays controlled without becoming too narrow.
Review criteria
The answer will be checked against 3 criteria. This makes the review step much more precise.
Structure and bans
Either the required blocks or the bans are missing. The shape or style of the result may get too loose.
Active review criteria
- Clarity
- Audience fit
- CTA strength
Ready prompt template
Copy and adaptYou are an experienced copywriter. Product: [product or service]. Audience: [segment]. Goal: [outcome]. Write the text in simple language, no longer than 140 words. Avoid overly formal or artificial phrasing. Required blocks: headline, main benefit, CTA. Restrictions: empty promises, unsubstantiated superlatives, mixing with jargon. At the end, self-check your answer against 4 criteria: clarity, brevity, audience fit, CTA strength.
Why it works
A length constraint keeps the answer from sprawling and forces the model to concentrate on what's important.
Tone helps adapt the text to the audience: simple, confident, friendly, or expert.
Restrictions cut off undesirable directions: clickbait, empty promises, unnecessary jargon.
Quality criteria transform the response from just text into a verifiable work product.
Practice
- Take one prompt for a Telegram post or ad.
- Add 3 constraints: length, tone, and a ban on undesirable wording.
- Write 3 quality criteria: clarity, CTA strength, audience fit.
- After the model responds, check it against those criteria and improve the prompt once.
Mini-project
Mini-project: controlled ad prompt
Create a repeatable prompt for one product or service that consistently produces a short and verifiable ad text.
Tasks
- Define the product, audience, and goal.
- Add 3 required blocks and 3 restrictions.
- Write 4 quality criteria.
- Run the prompt and evaluate the result against criteria.
Deliverables
- 1 controlled prompt template
- 1 AI response
- brief evaluation by criteria and an improved version
Checklist
Common mistakes
- adding too many constraints and stifling the prompt
- setting contradictory criteria
- listing only restrictions without required blocks
- using overly vague criteria like 'make it powerful'
Lesson FAQ
Will too many constraints make the result worse?
Not necessarily. If constraints are useful and don't contradict each other, the result usually improves. The issue isn't quantity but quality of constraints.
Should criteria be written at the end of the prompt?
Not mandatory, but a separate criteria block makes the prompt clearer for both the model and the human, especially if you want to build in a review step.
Check your knowledge
3 questions · Result saved in your browser
What will you gain from this lesson?
Which mistake is most common?
Will too many constraints make the result worse?