No AI Slop: evidence-led editing, packaging and evaluation
No AI Slop architecture: two response paths and one editorial checklist
Trace how the draft, mode choice and preservation constraints lead to either findings or a rewrite, and distinguish instructions from enforcement.
What you will learn
- The host provides execution; the skill provides a contract
- Detection exits before the rewrite path
- The checklist loops over meaning and style
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
- The host interprets instructions; the package does not contain a deterministic prose engine.
- Detection should exit without rewriting.
- Self-review is different from independent enforcement.
The host provides execution; the skill provides a contract
The assistant host loads the packaged instructions and receives the user’s draft. SKILL.md tells it to understand the draft’s point, audience and voice before acting. If needed context is missing, the instructions request clarification. There is no dedicated parser in this repository that deterministically classifies every sentence into a pattern.
The manifest describes how the package is presented and where skills are found. It does not implement the editing transformation. The Python builder belongs to distribution tooling; invoking it cannot simulate an assistant conversation. Keeping these components separate prevents a packaging success from being mistaken for a completed editing workflow.
Detection exits before the rewrite path
For a detection request, the specified result is a list of named patterns with quotations and brief repairs. The workflow stops there. An edit request instead makes minimal changes and checks the result against eval.md before returning the full edited draft and a change explanation. The two modes should not be merged into a single output that rewrites without being asked.
These branches are natural-language instructions interpreted by the model. The repository does not provide a runtime gate that blocks a response containing an authorship score or invented statistic. A consuming application that needs enforceable guarantees must add separate validation and a human review policy rather than assume the prose contract is executable control flow.
The checklist loops over meaning and style
eval.md asks whether the edit preserves facts, voice and recognizable cadence while removing named patterns. If a check fails, the instructions ask the model to repair the draft and check again. This is a self-review loop, not an independent evaluator with a recorded dataset, model configuration and measured success rate.
The architecture diagram therefore shows both response paths and the checklist only on the edit branch. Human acceptance remains outside the loop: a model marking its own response as compliant does not establish that a quoted fact survived or that the author recognizes the voice. Save original and revised drafts so those claims can be inspected.
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
Map the host, manifest, rules and checklist separately.
- 2
Identify the selected response mode.
- 3
Inspect the expected output contract for that branch.
- 4
Keep original and revised text for human comparison.
Copy-ready example
{
"detect": {
"rewrite": false,
"authorshipScore": false,
"output": [
"pattern",
"quote",
"short fix"
]
},
"edit": {
"selfCheck": "eval.md",
"output": [
"full draft",
"change explanation"
]
},
"runtimeEnforcementProvided": false
}Frequently asked questions
Is eval.md an executable test runner?
No. It is a checklist expressed in prose for the assistant to apply.
Does the package prevent forbidden outputs at runtime?
No separate runtime enforcement mechanism is established by the inspected files.
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
- No AI Slop / .codex-plugin/plugin.jsonSource checked 2026-09-14
- No AI Slop / scripts/build_plugin.pySource checked 2026-09-14