Evaluate and iterate — FACTS criteria

Section 10.2.9 : Evaluate and iterate — FACTS criteria

🎯 Learning objective

Developing a systematic method to evaluate AI response quality and iterate effectively until obtaining a usable result. Mastering the FACTS framework and iteration strategies to never again accept a mediocre response.

The trap of passive acceptance

The biggest risk with generative AI isn't that it's wrong — it's that you accept its response without critical evaluation. A good AI practitioner never takes the first response at face value.

Why this trap is so common:

• →AI produces fluent and confident responses — the "expert" tone creates an illusion of competence
• →Cognitive fatigue pushes us to accept the first response when it's "roughly correct"
• →Automation bias: we place excessive trust in automated systems
• →The cost of iteration seems high — but it's always lower than the cost of using a mediocre result

In reality, the difference between an average AI user and an expert AI user comes down to one word: iteration. The expert systematically evaluates and iterates 2-3 times. The average user takes the first response.

FACTS criteria: your evaluation framework

Fidelity — Actionability — Completeness — Tone — Structure

Scoring in practice:

• →Total score ≥ 40/50: the response is usable as-is
• →Score 30-39: a targeted iteration will suffice
• →Score < 30: reformulate the initial prompt — iteration won't be enough

The 5 iteration strategies

Strategy 1: Surgical iteration — Target the specific problem

The most effective and fastest. Precisely identify what's wrong and request a surgical correction.

Good response overall, but 3 adjustments needed:

1. The pricing section is too vague. Replace 
   generic ranges with concrete prices 
   for the French market in 2026:
   - Starter: [exact range]
   - Pro: [exact range]
   - Enterprise: [exact range]

2. The tone is too academic for my audience (SMBs, 
   non-technical executives). Reformulate paragraphs 
   3 and 5 in direct business language, no jargon.

3. The legal dimension is completely missing. 
   Add a paragraph about GDPR and compliance 
   obligations.

Keep everything else identical.

Strategy 2: Forced self-evaluation — Have the AI critique itself

Powerful technique when you're not sure what's wrong. The AI often identifies weaknesses you wouldn't have seen.

Reread your response above and rate it from 1 to 10 
on each FACTS criterion:

- F (Fidelity): /10 — is the information verifiable?
- A (Actionability): /10 — can someone act immediately?
- C (Completeness): /10 — are all aspects covered?
- T (Tone): /10 — appropriate for the target audience?
- S (Structure): /10 — clear and well-organized?

For each score < 8:
1. Explain the weakness precisely
2. Propose a concrete improvement
3. Rewrite the relevant section

Then produce the final version integrating all 
improvements.

Strategy 3: Confrontation iteration — Simulate critical external feedback

Ideal when the response seems "correct" but lacks depth. We simulate feedback from a demanding expert.

You show your response to 3 people:

1. A domain expert (15 years of experience) who says: 
   "This is superficial. A junior could write this."
2. The CEO who says: "Where are the numbers? 
   I can't make any decision with this."
3. A client who says: "Concretely, what do I do 
   Monday morning at 9 AM?"

Rewrite your response addressing each criticism:
- More technical depth (expert)
- More quantified data (CEO)
- More concrete step-by-step actions (client)

Strategy 4: Perspective shift — See the problem differently

When classic iteration isn't enough, change the angle of attack.

Your previous response approached the subject from 
the [PERSPECTIVE A] point of view. Now, rewrite it from 
the perspective of:

1. A skeptic who doubts every claim
2. A complete beginner with no context
3. A competitor looking for flaws

Then merge the 3 perspectives into a balanced response 
that anticipates objections.

Strategy 5: Subtraction iteration — Make it more concise

Often, the problem isn't missing content — it's too much. The superfluous drowns the essential.

Your response is [X] words. Rewrite it in [X/2] words maximum 
by removing:
- Everything obvious or already known to the audience
- Unnecessary transitions and filler phrases
- Redundant examples (keep the best of each category)
- Empty qualifiers ("very", "really", "quite")

Rule: each remaining sentence must provide information 
the reader didn't have before reading it.

The 4-step iteration workflow

Concrete iteration examples

Case 1: The too-vague report

Initial prompt: "Analyze SaaS market trends in France"

• →F: 5/10 (unsourced figures), A: 4/10 (no recommendations), C: 6/10 (B2C ignored), T: 8/10, S: 7/10
• →Score: 30/50 → Surgical iteration needed

Iteration:

Good foundation, but needs 3 major corrections:
1. FIDELITY: cite your sources for each figure. 
   If you don't have a source, indicate "estimate" 
   and give the range.
2. ACTIONABILITY: for each identified trend, 
   add "What this means for a French SaaS SMB: 
   [concrete action]".
3. COMPLETENESS: you ignored the B2C segment and the 
   vertical market. Add a section on each.

Case 2: The too-gentle code review

Initial prompt: "Do a code review of this Python file"

• →F: 8/10, A: 5/10 (vague suggestions), C: 3/10 (security absent), T: 9/10, S: 7/10
• →Score: 32/50 → Confrontation + surgical

Iteration:

Your review is too lenient. A senior lead engineer 
would read it and say: "You didn't check security 
and your suggestions don't contain corrected code."

Redo your review adding:
1. Complete SECURITY section (SQL injection, XSS, auth, 
   exposed sensitive data)
2. For EACH identified problem: the current code AND 
   the corrected code, not just a description
3. A summary table with severity and priority

Practical exercise: The FACTS challenge

Duration: 20 minutes

1. →Take a recent AI response you've used in your work
2. →Rate it on each FACTS criterion (F, A, C, T, S) from 1 to 10
3. →For each score < 8, write a targeted iteration prompt
4. →Send the iterations and compare V1 and V2
5. →Does V2 have a score ≥ 40/50? If not, try the confrontation strategy