# Skill: Build a structured AI prompt, run targeted iterations, and apply the pre-publish review checklist

> Lets an AI agent construct a five-element prompt, guide the user through 2-4 productive edit
> passes, and run a five-question review before any content is used; applies to any text-generation
> task where the result must meet a quality standard rather than "good enough on first pass".

Source: sodigi·learn - neuroslop/review-i-iteratsii · https://sodigi.io/learn/neuroslop/review-i-iteratsii

## When to use
- When building a prompt for any piece of content that matters: post, email, proposal, article, announcement.
- When the user has a draft and wants to know how to improve it without rewriting from scratch.
- When an AI-generated result looks polished but is generically correct and would pass the replaceability test.
- When a text is "almost done" and needs a final check before publication or delivery.

## Core rules
- The first answer is a draft and an invitation to talk, not a final product. Quality work takes 2-4 passes.
- Model smoothness (grammar, transitions, clean paragraphs) is a default property, not a signal of quality. Judge what it says, not how it is written.
- Productive iteration names a spot, a problem, and a direction. "Redo it" and "make it better" are not edits; they just reshuffle the same average.
- The five prompt elements each prevent a specific type of slop: role shifts from "internet average" to a narrower expertise register; context supplies facts the model cannot invent; task defines the reader outcome, not just the topic; constraints cut stock reflexes; example communicates voice where words cannot.
- Iterate to a defined standard, not to exhaustion. Determine the "done" criterion before starting, not after you get tired.
- Every fact that can be checked gets checked before publication. The model produces invented and real facts with identical smoothness and confidence.
- The goal is "I decided and edited faster" - not "AI did it for me". The operator's context, position, and standard must remain in the work.

## Procedure
1. BEFORE prompting, spend two minutes collecting: audience (who they are, what they know, what they want), your facts (numbers, cases, specific data), the reader outcome (what should they do or feel after reading), tone (adjectives plus a sample sentence or past piece they can match), and constraints (length limit, things to avoid).
2. BUILD the prompt using all five elements:
   - Role: "You are [editor / marketer / instructional designer...]"
   - Context: who you are, your product, the audience, your specific facts and data
   - Task: what to produce, and what the reader should do or feel afterward
   - Constraints: length, tone, list of things not to do (no corporate-speak, no invented facts, no exclamation marks)
   - Example: paste a fragment whose voice you want to match
   - Close with: "First show me a plan in 3-4 bullet points, then wait for my OK."
3. REVIEW the plan before generation: correct direction, emphasis, and structure at this stage rather than after a full draft.
4. RECEIVE the draft. Read it for what it says, not for grammar. Apply the replaceability test immediately.
5. ITERATE with targeted edits. Each edit must name: (a) the spot ("first paragraph"), (b) the problem ("it is filler that repeats the subtitle"), (c) the direction ("cut it and open with the pain: quarterly report eats three evenings"). Run 2-4 such edits.
6. RUN the five-question review checklist:
   a. FACTS: every number, name, date, quote, link verified against a primary source?
   b. TONE: read aloud - does it sound like you, or like corporate boilerplate? Rewrite any spot where you would never say it that way.
   c. ANCHOR DETAIL: is there at least one detail that could only be written about this specific product or person? If not, add one.
   d. REPLACEABILITY TEST: substitute a competitor. Does the text still read as correct? If yes, return to context and add specific material.
   e. ADJECTIVES: mentally strike out all evaluative adjectives ("unique", "innovative", "best", "high-quality"). Does the text keep its meaning? If adjectives were load-bearing, the content beneath them is empty.
7. PUBLISH once all five checklist questions pass. If one fails, run one more targeted iteration to fix that specific item.

## Ready-to-use prompt
```
Role: you are [an editor / marketer / instructional designer - fill in the right role].

Context: I am [who you are and what your product is]. I'm writing for [audience: who they are, what they know, what they want]. My facts and data: [numbers, cases, observations - all specific material you have].

Task: [what to produce]. After reading, the reader should [specific action or takeaway].

Constraints: length [how long], tone [what kind], do not use [cliches, corporate-speak, exclamation marks, etc.], do not invent facts - use only the data above.

Tone example - here is a text whose voice I want to match:
[paste the fragment]

First show me a plan in 3-4 bullet points, then wait for my OK before writing the full text.
```

## Pitfalls
- Do not copy and publish the first answer. The first answer is a draft even when it looks finished.
- Do not use vague feedback ("make it more interesting", "redo it", "write it better"). Each iteration must name a spot, a problem, and a direction.
- Do not skip the plan step. Asking for a plan before the full draft lets you correct direction before the model writes a thousand words the wrong way.
- Do not iterate to the point of fatigue and call the result "done". Fatigue is the worst editor; define "done" with a checklist criterion before you start.
- Do not treat the review checklist as optional for quick tasks. A two-minute check on a social post is faster than managing the fallout from a published error.
- Do not skip the example element of the prompt when tone matters. One sample paragraph communicates voice more precisely than any description of tone in words.
- Do not confuse the number of iterations with the quality of iterations. Five vague passes produce worse results than two targeted ones.