# Skill: Diagnose which of the five operator error scenarios is producing slop and prescribe the fix

> Lets an AI agent identify the specific operator mistake causing low-quality AI output and return
> a targeted, scenario-specific remedy; applies whenever a user reports that AI "isn't working" or
> keeps producing mediocre results despite trying.

Source: sodigi·learn - neuroslop/pochemu-lyudi-delayut-slop · https://sodigi.io/learn/neuroslop/pochemu-lyudi-delayut-slop

## When to use
- When a user is frustrated that AI output is generic, "feels fine but something's off", or not working.
- When reviewing a user's workflow to find where quality is leaking before content reaches publication.
- When helping a user improve their prompting practice without telling them to "just write better prompts".
- When a user asks why AI keeps producing the same average-sounding text no matter what they try.

## Core rules
- A language model with no specific input returns the weighted average of its training data; that average is slop by definition.
- All five operator scenarios are properties of the human, not the model. Same model, different operators, different quality classes.
- AI is a competence mirror: it amplifies what is fed to it. Feeding emptiness returns amplified emptiness faster.
- The core AI skill is knowing what you want to get - audience, goal, facts, quality criterion - before composing the prompt.
- Scenario 1 (volume greed): generate only as much as you will seriously edit. Twenty average posts produce fewer results than one strong one.
- Scenario 2 (no context): anything absent from the prompt gets replaced by the internet average; your facts, audience, and position must be in the input.
- Scenario 3 (no taste): if you cannot name what is strong in the result in three specific sentences, nothing likely is; compare against the best in the niche, not against a blank page.
- Scenario 4 (no fact-checking): model confidence is not correlated with accuracy; every checkable fact must be verified before publication.
- Scenario 5 (no iterations): the first answer is a draft and an invitation to talk; quality results typically appear on iterations 2 through 4.

## Procedure
1. Ask the user (or infer from context) five diagnostic questions: (a) How many pieces were requested at once? (b) How long was the prompt - did it include audience, facts, and goal? (c) Can they name what was specifically strong in the last result they accepted? (d) Were numbers and links in the last published text verified? (e) How many iterations did the last task receive?
2. Map each uncomfortable answer to a scenario: volume = Scenario 1; vague prompt = Scenario 2; "it seemed fine" = Scenario 3; unchecked facts = Scenario 4; one-pass copy-paste = Scenario 5.
3. Prescribe the scenario-specific fix:
   - Scenario 1: set a batch limit to what can be seriously edited in the same session (typically 1-3 pieces).
   - Scenario 2: rebuild the prompt with audience, specific facts, goal, and tone before next generation.
   - Scenario 3: place the result beside the strongest competitor example and name three concrete differences.
   - Scenario 4: extract every number, name, date, and link and verify each one against a primary source.
   - Scenario 5: run at least two more targeted iterations naming spot, problem, and direction for each edit.
4. If multiple scenarios are active, address Scenario 2 (no context) first - it is the highest-leverage fix.
5. Confirm the fix worked by re-running the slop checklist on the revised result.

## Ready-to-use prompt
```
You are an AI workflow diagnostician. A user is getting low-quality AI output. Run the five-scenario diagnosis below.

Ask the user (or analyze the information provided) to answer these five questions:
1. How many pieces of content did you request in the last session? How many did you edit before using?
2. Paste your last prompt. Did it include: the audience, your specific facts/numbers, the goal, and the tone?
3. Describe in three specific sentences what was strong about the last result you accepted. (If you cannot, say so.)
4. Did you verify every number, name, date, and link in the last text you published?
5. How many messages did the last task take: one pass, or two to four iterations?

Based on the answers, identify which scenario(s) are active:
- Scenario 1 VOLUME GREED: batch too large to edit seriously
- Scenario 2 NO CONTEXT: prompt lacked audience / facts / goal / tone
- Scenario 3 NO TASTE: operator cannot name what is specifically strong in the result
- Scenario 4 NO FACT-CHECKING: model output accepted without verification
- Scenario 5 NO ITERATIONS: first answer copied and published without refinement

For each active scenario output: SCENARIO NAME | what is happening | one concrete fix to apply right now.
Prioritize Scenario 2 if active - it is the highest-leverage repair.
```

## Pitfalls
- Do not blame the model when diagnosing poor output; the correct intervention point is always the operator's inputs and process.
- Do not prescribe "write better prompts" as the fix for Scenario 3 (no taste); taste is trained by deliberate dissection of strong examples, not by prompt refinement alone.
- Do not treat Scenario 4 as rare or only relevant to legal/medical contexts; the model invents facts confidently across all domains.
- Do not allow the user to fix Scenario 1 by editing faster rather than generating less; speed editing produces slop just as surely as no editing.
- Do not conflate Scenario 5 with "keep iterating until perfect"; the goal is a defined quality standard reached in 2-4 passes, not endless refinement.
- Do not skip the self-diagnosis: the most useful version of this skill is applied to the user's own workflow, not only to other people's output.