No AI Slop: evidence-led editing, packaging and evaluation
Try No AI Slop with a fact-preserving release note exercise
Use one synthetic paragraph to test detection and minimal editing while protecting dates, technical details and deliberate uncertainty.
What you will learn
- Define the facts before requesting an edit
- Start with detection so the evidence is visible
- Make the acceptance decision yourself
Before you start
- A draft whose facts can be checked
- Basic understanding of assistant instructions and plugin scope
Inspect pattern evidence, preserve meaning and distinguish package checks from unmeasured editing outcomes.
Key takeaways
- Protect facts and uncertainty before editing.
- Detection and rewriting need separate acceptance checks.
- A repaired answer should still be recorded as an initial failure.
Define the facts before requesting an edit
Use a disposable conversation and a reviewed copy of the skill. Provide a synthetic release note that contains one concrete feature, one date and an uncertainty statement. Mark those details as protected. The goal is to observe whether a response removes an empty flourish while preserving information, not to reward the shortest possible answer.
Our example says draft search shipped on Tuesday and mobile behavior has not been tested. Those are invented exercise inputs, not claims about this project. An acceptable rewrite retains the feature, Tuesday and the untested mobile boundary; it cannot turn an unfinished check into a promise of mobile support.
Start with detection so the evidence is visible
Ask for detection without rewriting. A useful response quotes the sentence it flags, names the matching pattern and suggests a small repair. Reject an answer that supplies an AI-likelihood percentage, rewrites the whole paragraph or identifies ordinary technical language as an authorship signal. Those outputs violate the intended detection contract.
Then request editing in a separate step using the same source draft and protected facts. Compare the complete revised text with the original. Ask whether each deletion removes redundancy or removes meaning. A source-specific detail is more useful than a polished replacement that could describe any product.
Make the acceptance decision yourself
The editing output should include the full draft and a concise change explanation. Check that the explanation matches the actual modifications, including any reordering. If the assistant adds a statistic or removes genuine uncertainty, restore the original and record the failure before trying another configuration. Do not silently fix the response and count it as a clean pass.
This is a reader exercise, not an experiment executed in a live assistant during our review. Keep installation and model version in your trial record. A new clean conversation is the simplest baseline; an existing conversation may already contain the candidate rules and contaminate the comparison.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Prepare the synthetic note and its protected facts.
- 2
Request findings only and inspect quoted evidence.
- 3
Request the minimal edit separately.
- 4
Compare the complete text and change explanation.
Copy-ready example
Synthetic draft: Draft search shipped Tuesday, highlighting our commitment to innovation. We have not tested mobile behavior.
First detect named patterns without rewriting or guessing authorship.
For a later edit, retain: draft search; Tuesday; mobile not tested.Frequently asked questions
Should shorter always score higher?
No. A short answer that drops an untested condition or technical fact fails the exercise.
Was this prompt tested with a live model here?
No. It is a proposed reader exercise with explicit acceptance criteria.
Sources
- No AI Slop / skills/no-ai-slop/SKILL.mdSource checked 2026-09-14
- No AI Slop / skills/no-ai-slop/eval.mdSource checked 2026-09-14