No AI Slop: evidence-led editing, packaging and evaluation
Build an evidence-first editing worksheet for No AI Slop
Design a small review aid that anchors findings to draft spans and protects facts, without turning stylistic preferences into automatic rewriting.
What you will learn
- Make every proposed edit point back to the source
- Separate mechanical checks from editorial decisions
- Define failure cases before connecting a model
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
- Anchor findings to a specific draft revision and span.
- Mechanical validity does not establish editorial quality.
- Applying an edit should require an explicit decision.
Make every proposed edit point back to the source
A useful extension is a worksheet containing the original passage, named pattern, proposed edit, reason and approval state. It should also record protected facts and quotations. This is an editorial project proposal, not a feature shipped in the repository or a claim about its roadmap. Begin with synthetic drafts and manually entered findings.
Store the draft revision and a digest so a finding cannot silently apply to changed text. A phrase repeated twice needs a location, not just a search string. If a later draft no longer matches the cited span, mark the finding stale and request review rather than replacing the first similar phrase.
Separate mechanical checks from editorial decisions
A small offline program can verify that a quoted span exists, a protected date remains and an accepted edit targets the recorded revision. Those mechanical checks cannot determine whether irony survived or whether a sentence should be shorter. Represent those judgments as pending human decisions rather than fabricating a confidence score.
Use a side-by-side text view and static annotations before adding animation. An accessible SVG can explain the relation among original text, findings and approved edits. Three-dimensional presentation offers little benefit for this task; the source text and revision history must remain readable without an interactive scene.
Define failure cases before connecting a model
Test a missing quotation, duplicated phrase, stale revision, protected number change and a finding that asks for authorship inference. The worksheet should refuse automatic application in each ambiguous or disallowed case while preserving the original. These are proposed extension tests, separate from the seven package-validation cases we executed.
A later model integration can propose findings, but applying them remains an explicit action. Export the reviewed draft and a change record together so a downstream editor can inspect the reasoning. A green package build or a model’s self-check must never be substituted for approval of the actual text change.
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
Define original text, finding and protected-fact records.
- 2
Reject stale or ambiguous text anchors.
- 3
Test mechanical invariants with synthetic drafts.
- 4
Add model suggestions only after the review path works.
Copy-ready example
{
"proposal": true,
"draftRevision": "synthetic-v1",
"finding": {
"spanStart": 0,
"spanEnd": 12,
"pattern": "example",
"approved": false
},
"protectedFacts": [
"Tuesday"
],
"applyAutomatically": false
}Frequently asked questions
Is this worksheet already part of the project?
No. It is a proposed learning extension.
Can a digest determine whether the edit is good?
No. It identifies the source revision; quality still requires contextual review.
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 / scripts/build_plugin.pySource checked 2026-09-14