Ben AI built a De-Slop skill: a Claude skill you run on any AI output before it goes out, to catch "slop" the way spell-check catches typos. His argument is that good prompts and context still drift, and your own quality bar slips when you are tired or rushed, so the real fix is a consistent gate that holds every output to a defined standard. This note covers how the skill works and compares it to my own anti-AI-writing guidelines.
One caveat up front: the skill file is behind an email opt-in (it redirects to a Kit landing page), so I could not pull the raw markdown. The design below is reconstructed from the video, not the literal file.
How the skill works
The skill runs on a finished AI output and grades it in two layers:
- Universal slop check (objective). The agreed-upon tells: em-dashes, AI vocabulary and writing patterns, internal contradictions, unreadable writing, plus a fact-check of claims against authoritative web sources.
- Company / personal slop check (subjective). The part you personalize: tone of voice, brand identity (including visual guidelines for visual assets), alignment with your strategy and facts, and whether the piece actually does its job.
The process:
- Detect the output type (marketing post, support reply, internal doc) and pick which checks apply.
- Run each check against a reference file. Each check (AI writing tells, factual accuracy, voice, brand) has its own guidelines doc. The voice doc, for example, says how the brand should and should not sound ("no guru energy, no big claims without backing").
- Grade with three sub-agents against the universal and company criteria, kept separate to reduce bias.
- Return a verdict: "good to go", "good to go but fix a few things", or "not ready", with a table of what is off, the source backing each flag, and a suggested rewrite. You choose which fixes to apply.
- Log the run to a De-Slop Log, so recurring slop patterns and usage show up over time.
Slop, as Ben frames it, is subjective and tied to your quality bar. The skill's job is to enforce that bar consistently even when you cannot be bothered to.
For business people
The bigger problem is institutional. Every employee has a different quality bar, and each person's AI just agrees with them rather than enforcing the company's standard, so more AI use can mean more "work slop" out the door. Ben cites an a16z piece on institutional vs individual AI: AI made individuals more productive, but no company became 10x more valuable, because speed without a shared direction or quality benchmark creates chaos, not value.
Run as a team gate, the skill gives everyone the same bar before anything ships, and the log makes the company's recurring weak spots visible. It pairs with a shared second brain for context (see Setting Up an AI OS: 6 Mistakes and Their Fixes), but context alone does not guarantee the output meets the bar, which is the gap this check fills. Some companies go further and ban AI for anything public-facing; a shared de-slop gate is the lighter alternative.
For technical people
How it is built and how you would wire it in:
- Reference files per check. Each rule set (AI writing tells, factual accuracy, voice, brand, visual) is its own markdown guideline the grader reads. Swapping or extending a check means editing a file, not the logic.
- Output-type routing. The skill first classifies the asset, then runs only the relevant checks.
- Sub-agent grading. Three sub-agents grade in parallel for less bias, then a verdict is composed with issues, sources, and suggested fixes.
- Composable. It can be the last step inside another skill. Ben bolts it onto the end of his LinkedIn-post skill so the draft is auto-checked before he sees it.
- Personalization. The shipped skill has blanks (voice, brand, facts) you fill, either by asking Claude to draft them or via their paid "De-Slop Builder" skill.
- Logging. Each run appends to a De-Slop Log in the second brain (or in the skill folder) for pattern detection.
Compared to my anti-AI-writing guidelines
My anti-ai-writing-guidelines.md and this skill attack the same problem from different ends.
| Dimension | My guidelines | Ben's De-Slop skill |
|---|---|---|
| Form | A static reference I read and apply by hand | A runnable skill that gates output before it ships |
| Universal style tells | Deep taxonomy: vocabulary blacklist, the "-ing significance" habit, rule of three, "not X but Y", false ranges, em-dashes, promotional tone, weasel words | Same idea but lighter: AI words and patterns, em-dashes, contradictions, unreadable writing |
| Voice / brand layer | Generic ("adopt my voice"); my real voice lives in separate files | Built-in, personalized voice + brand + strategy + facts, per output type |
| Fact-checking | Not covered, style only | Systematic claim-check against sources |
| Output | A self-check list | Verdict (good / fix / not ready) + issue table + sources + rewrite, graded by sub-agents |
| Feedback loop | None | A log of every run to surface recurring patterns |
| Scope | Me, manually | Me and a whole team, enforced consistently |
The short version: my guidelines are the stronger universal reference, and they would slot straight in as the skill's "AI writing tells" file. What the skill adds is everything around that doc, the enforcement, the personalized voice layer, the fact-check, the verdict, and the log.
What I would take from it
- Turn my guidelines into a skill, not a doc. The honest problem Ben names is mine too: the bar drops when I am rushed. A gate that runs automatically beats a doc I sometimes forget.
- Wire in my existing voice files (
my-linkedin-style.json,my-email-style.json,my-rfp-style.md) as the personalized layer, so the check knows my voice, not just generic AI tells. - Add a fact-check and a verdict. My guidelines are style-only; a claim-check plus a clear good / fix / not-ready call would make it operational.
- Bolt it onto the end of my writing skills, especially the LinkedIn-post flow, so every draft is pre-checked.
Further reading
- Ben AI video: How to De-Slop Every AI Output Forever (With 1 Skill)
- a16z: Institutional AI vs Individual AI
- The free skill (email opt-in): Ben AI De-Slop skill
- Related: Setting Up an AI OS: 6 Mistakes and Their Fixes