Start with Humanizer if you want editing rules that explicitly adapt to a writing sample. Choose Stop Slop if you want a stricter house style with firm rules against stock phrases and sentence patterns. If your current editing prompt already gives you good results, you may not need either: in our small English comparison, a prompt without a skill removed the same obvious marketing language.
Both are open-source instructions for an AI agent to follow. Their rules offer a useful basis for choosing, but our five runs did not establish a quality winner.
Disclosure: I build StashBase, which appears in the writing-library workflow below. StashBase was not involved in these tests.
How Humanizer and Stop Slop differ
The clearest difference is how much freedom the rules leave for the writer’s own style.
Humanizer 3.0.0 tells the agent to identify patterns, rewrite, check for remaining problems and factual changes, then produce the final version. A supplied writing sample takes priority over its pattern rules. That includes punctuation: deliberate dashes in your sample can stay. Its instructions also distinguish personal writing from factual reference prose.
Stop Slop sets firmer constraints: remove adverbs and em dashes, use active voice, and avoid formulaic structures. It includes a self-rating checklist for directness, rhythm, trust, authenticity and density. Those ratings guide revision; we did not use them as independent quality measurements.
| Your preference | Where to start | What to check in the result |
|---|---|---|
| Preserve the rhythm of your own essays | Humanizer with a representative sample | Whether the result actually resembles that sample |
| Enforce a consistent, restrictive editorial style | Stop Slop | Whether a rule removes a useful qualification or deliberate construction |
| Clean up occasional drafts | Your existing editing prompt | Whether a reusable skill adds anything you keep having to request |
These are recommendations from the documented rules, not rankings from a large benchmark. You can edit either rule set to suit your publication. Both repositories use the MIT license: Humanizer, Stop Slop.
For setup, follow each project’s README. Our earlier Humanizer walkthrough covers installation and a separate test of version 2.11.2; this comparison uses 3.0.0.
The same English draft, edited by both
Both skills removed the conspicuous marketing language and retained the factual content in this example. The ordinary editing prompt did much the same.
We constructed a short blog draft about a fictional research team. The numbers and events below are test material, not results from a real study. The draft deliberately mixes promotional filler with concrete facts and an author’s reservation:
I’m excited to share a game-changing update from our four-week trial of a shared interview notebook. Eight researchers used it to record 17 interviews, and the results speak volumes. It’s not just about collecting notes; it’s about unlocking a whole new way to collaborate.
Before the trial, I expected everyone to abandon their private documents. In reality, five researchers kept drafting privately and copied their finished notes into the shared notebook. Three wrote there directly. Two interview records lacked consent status, so we held those back from the public summary. We did not measure time saved.
The trial ended on 28 August. We’ll keep the shared notebook for another month, but I’m not ready to require it for first drafts. This is a powerful reminder that the future of research is connected, collaborative, and human.
Each run received the same draft and a list of facts to preserve, including the missing consent status, the absence of a time-savings measurement and the author’s reluctance to mandate first drafts. We requested only the final text.
Humanizer returned:
We ran a four-week trial of a shared interview notebook. Eight researchers used it to record 17 interviews.
Before the trial, I expected everyone to abandon their private documents. In reality, five researchers kept drafting privately and copied their finished notes into the shared notebook. Three wrote there directly. Two interview records lacked consent status, so we held those back from the public summary. We did not measure time saved.
The trial ended on 28 August. We’ll keep the shared notebook for another month, but I’m not ready to require it for first drafts.
Stop Slop returned the same text with just “In reality,” removed. The no-skill control used “We recently completed” instead of “We ran” and otherwise matched Humanizer’s output.
| Condition | Words | Difference from Humanizer’s first result |
|---|---|---|
| Original draft | 135 | Includes the promotional opening and closing |
| Humanizer | 94 | Reference result shown above |
| Stop Slop | 92 | Removes “In reality,” |
| Humanizer with a writing sample | 94 | Identical text |
| Stop Slop with a writing sample | 94 | Identical text |
| No skill, same editing request | 95 | Changes the opening to “We recently completed” |
All five outputs retained the supplied counts, dates, limitations and decision. None added a factual detail. All dropped the grand conclusion about the future of research. Shorter text alone was not the success criterion: keeping the consent restriction and the unmeasured outcome mattered more.
This easy example gave us no reason to prefer one skill’s output. It also showed that the model could perform this cleanup with the common editing request and fact constraints alone.
Did a writing sample make the result more personal?
It did not produce a meaningful change in these runs. Humanizer’s output stayed identical; Stop Slop added back the two words it had removed.
For the sample conditions, both received this separate, synthetic passage:
I moved my reading notes into a single folder last winter. I liked the idea more than the first week of using it. Naming the files took longer than I expected, and I kept opening the old folder out of habit.
The small change that stuck was a date at the start of each filename — I could find Tuesday’s notes without remembering what I’d called them. I still leave unfinished notes on my desktop. That probably says more about me than the system.
We asked the model to match its voice and rhythm where compatible with the editing rules, and to use it for style only. The draft and factual constraints stayed unchanged.
The result does not establish that voice matching fails. We used one short draft with much of its useful prose already written plainly, one sample and one run per condition. It does mean we cannot claim that adding a sample improved this draft.
Try a passage from your own work before adopting either skill. Keep the original beside the rewrite and check whether your opinion, uncertainty and meaningful word choices survive. If the rewrite sounds interchangeable with anyone else’s, identify a specific passage in your sample that shows the difference you want.
Keep writing samples and factual sources ready
A few pasted paragraphs are enough to try these skills. A reusable library becomes useful when you repeatedly need to find relevant examples of your own writing, retrieve interview notes or check a claim against an older document.
In StashBase, you can keep your published pieces, drafts and source notes in local folders. File preparation makes supported documents and recordings searchable; Search by meaning helps locate related material when you do not remember its wording. A Wiki can organize recurring topics with links back to the sources. See the search guide and Wiki setup.
A practical workflow is to retrieve a relevant writing sample and the source material first, then give them to your agent with the draft and your chosen editing skill:
Find two of my published pieces with a similar audience and tone.
Show the passages and source links before editing this draft.After reviewing those selections:
Use Humanizer to edit this draft with the approved samples for voice.
Use the source notes for facts. Flag missing details; do not invent them.Connect your agent to StashBase to retrieve material from the same library. This is a proposed workflow, not a result from the comparison above. Its value is keeping useful evidence and examples available across writing projects; our tests did not measure whether StashBase improves prose.
Test conditions
We ran five fresh sessions on September 10, 2026, in the Asia/Shanghai timezone, using Claude Code 2.1.220. The requested model was sonnet; the editing model reported in every response was claude-sonnet-5, with effort set to medium.
We supplied the skill text as system instructions, including all three reference files for Stop Slop. Tools, external connectors and customizations were disabled. This tests the supplied instructions, not automatic skill loading. Humanizer was version 3.0.0 at commit 9862685; Stop Slop was at 8da1f03. The no-skill control retained the common editing request and fact constraints but omitted the skill text.
Each condition ran once. The draft and sample were constructed for this test. Word counts split the text on whitespace. We reviewed factual retention and visible changes, with no detector scores or blind preference study. The results apply to these inputs and settings.