The Listing Audit Nobody Runs Until the Account Is Already at Risk
Cross-border sellers spend real money on copy and almost none of it on copy governance. A single Amazon bullet can carry a compliance promise that triggers a suppression; a German product page can mix “Du” and “Sie” across three translated variants and quietly erode conversion in the DACH market; a Klaviyo flow written by four different freelancers can drift into the same hollow AI cadence until open rates sag and nobody can name why. The bottleneck was never writing — it’s the fact that nobody lints the output. slop-grader, an open-source tool from Lukas, is the first thing I’ve seen that treats text compliance the way we treat code: as a ruleset you run repeatedly, not a vibe you hope holds. That’s why it matters to operators, even if the tool itself is still rough.
What It Actually Does — and What It Refuses to Be
Strip away the Product Hunt framing and slop-grader is a rules engine for prose. You define checks as plain-language questions, and the tool evaluates them either line by line (“Does this line make a promise that requires a legal disclaimer?”) or across the whole document (“Does the opening earn the reader’s next 30 seconds?”). It ships with built-in rulesets for English grammar, German grammar, and AI filler detection, and you can author custom rules for your own workflow — SEO checks, legal clauses, tone-of-address consistency.
The mechanics matter for cost planning. It runs on Jev, which Lukas describes as a “System One” model — not an LLM — specialized in answering structured questions. The claim is that checking a document takes seconds and costs less than a cent. It outputs a list of flagged lines plus instructions you can paste directly into an AI agent to fix the document, which Lukas confirmed in the thread when asked whether the output was agent-ready.
Setup is not consumer-grade. You need Node.js installed locally and an account with either TypeSafe or OpenRouter, and — this is the line most sellers will skim past — text is evaluated on an external AI server. For a brand running unpublished Q4 creative or a legal review of marketplace claims, that’s a procurement question, not a footnote.
Why Amazon sellers should care more than Shopify ones
Shopify gives you latitude. Amazon does not. Your Amazon Seller Central listing copy is subject to automated and manual policy review, and the failure mode of a bad claim isn’t a lower CTR — it’s a suppressed listing, a hijacked buy box window while you appeal, and inventory aging in FBA. A line-by-line rule like “Does this line make a promise that requires a legal disclaimer?” is exactly the check that catches “FDA-approved,” “cures,” or “guaranteed results” before a competitor reports you. Shopify merchants writing DTC landing pages can absorb a soft claim; Amazon sellers cannot. If you sell supplements, topicals, or anything with a health adjacency, this category of tooling is closer to insurance than to copywriting.
How It Differs From the Stack You Already Pay For
Let’s be honest about the incumbents, because “AI writing tool with rules” is a crowded shelf.
- Grammarly owns grammar and tone. It’s excellent at the sentence level and useless at the workflow level — it doesn’t know your legal clause library, and it won’t enforce “Du” vs. “Sie” consistency across a German catalog.
- Jasper and Copy.ai are generation-first. They’ll produce the copy; they won’t audit copy that came from a freelancer, an agency, or a translated marketplace feed.
- Helium 10 and similar Amazon tooling will flag restricted keywords and indexing gaps, but the rules are theirs, not yours. You can’t write “Does this line make a promise that requires a legal disclaimer?” and have it run across your Etsy descriptions too.
- Translation QA layers like Gengo or agency-side review catch language errors but rarely catch tone-address drift or AI filler, because that’s not what you hired them for.
The differentiator Lukas keeps returning to in the thread is that the real power is custom rules for your use case, and that rules are evaluated independently against every line — cheap and fast enough that you can run dozens without watching a spinner. Gal Dayan made the sharpest observation in the comments: the Du/Sie example proves this isn’t an AI-filler detector, it’s “closer to a linter for anything you can phrase as a yes/no question about a line.” That reframe is the whole product. If you can phrase the check, you can automate the check.
The gap between “useful lint” and “lint you ignore”
Dayan also asked the question that determines whether this tool survives contact with a real team: when a rule fires, do you get just the line number and rule text, or does it explain why it thinks the rule was violated? Lukas’s answer is honest and, to my mind, the biggest limitation: it does not give you the reason. His workaround is to write separate rules for each check so the AI agent can infer the problem downstream.
That’s a real tradeoff. In a code linter, a bare rule ID is fine because the rule is deterministic and documented. In prose, a flagged line with no rationale means a human has to re-read the rule, re-read the line, and reconstruct the logic — every time. For a solo operator running ten rules, fine. For an agency running forty rules across twenty client catalogs, that’s the difference between a tool people trust and a tool people mute.
What Cross-Border Sellers Should Borrow From This
You don’t have to adopt slop-grader to steal its operating model. The transferable idea is treating brand voice and compliance as a versioned ruleset, not a style guide PDF that nobody opens.
Here’s how I’d map it to a real cross-border stack:
Marketplace listing hygiene. Build a ruleset per marketplace. Amazon: “Does this bullet make a comparative claim against a named competitor?” (risky). Etsy: “Does this description use a handmade or vintage claim we can’t substantiate?” TikTok Shop: “Does this script promise a result in a timeframe we can’t support?” These are yes/no questions about a line. That’s the entire format.
Localization tone locks. The Du/Sie example generalizes. In Japanese, are you consistent on keigo level across the PDP and the post-purchase email? In French, tu vs. vous? In Spanish, regional variants across your LatAm vs. Spain storefronts? A translated catalog assembled from multiple vendors will drift. A rule catches it in seconds.
Email and lifecycle copy. Klaviyo flows accumulate edits from everyone. A “Does this line sound like it was written by a human who has used the product?” rule is crude but effective at catching the AI filler that’s quietly homogenizing DTC email. Run it before every major campaign, not after.
Agent-ready output. The flagged-lines-plus-instructions format is the part I’d copy even without the tool. Whatever your QA process is, make its output pasteable into an agent. If your compliance review produces a PDF, it’s dead weight. If it produces instructions a Claude or GPT agent can execute against the source doc, you’ve compressed a two-day review into an afternoon.
Where the math breaks
The “less than a cent per document” claim is attractive, but it’s a marginal cost, not a total cost. Add the Node.js setup, the TypeSafe or OpenRouter account, the external-server data exposure, and — the real expense — the labor to author and maintain a ruleset that’s actually good. A bad ruleset produces false positives, and false positives are how tools die inside organizations. Budget the rule-authoring time as the dominant cost, not the inference.
Where My Judgment Says It Falls Short
Three things I’d push back on.
First, the missing rationale. As covered above, no “why” means every flag costs a human a re-read. That’s survivable at small scale and corrosive at large scale.
Second, no strictness dial. Amelia raised exactly this: a strictness setting so users can choose between a quick cleanup and a deeper review. It’s a good instinct and it’s not in the product. Without it, you’re forced to either run a heavy ruleset constantly or maintain two rulesets manually.
Third, the data-handling posture. “Text is evaluated on an external AI server” is stated plainly, which I respect, but for a brand sending unreleased product copy, legal-reviewed claims, or customer-adjacent messaging through a third party, that’s a conversation with legal before it’s a conversation with engineering. This is not a blocker for a solo seller; it is a blocker for an enterprise brand with a DPA requirement.
Fourth — and this is structural, not a bug — the “build custom rules” answer is doing too much work. Lukas answers the buzzword question and the rationale question both with “create custom rules.” That’s true and it’s also a way of saying the tool ships you a hammer and a pile of nails. The value is entirely a function of how good your rules are, and there’s a skill file in the GitHub repo to help author them — but rule quality is your problem, not the tool’s. That’s fine for operators who like building systems. It’s a wall for operators who want a product.
What I’d Watch / Test Next
This week, before you install anything, do the cheap version of the exercise.
Pick your highest-risk listing — the one with a health claim, a comparative claim, or a translated German or Japanese page. Write five yes/no questions about individual lines, in plain language, the way you’d ask a junior copywriter to check them. Then run those five questions manually against the copy. You’ll learn two things fast: whether your rules are specific enough to fire consistently, and whether your current copy survives them. Most sellers will find at least one line they’d rather not have in front of a marketplace reviewer.
Then, if the manual pass proves useful, install slop-grader, wire up the TypeSafe or OpenRouter account, and automate those five rules. Watch three things over the next month: false-positive rate (if it’s above roughly one in five, your rules are too vague, not the model), whether the flagged-lines output actually gets pasted into an agent or just gets read and ignored, and whether your legal or brand team objects to the external-server data path. If all three pass, expand the ruleset to your email flows and your Etsy or TikTok Shop descriptions. If any fail, you’ve still walked away with the real asset — a written, testable definition of what “acceptable copy” means for your brand. That’s worth more than the tool.
