Show HN: Detecting AI slop with regex and Stephen King

Notifications You must be signed in to change notification settings

Fork 0

Star 2

BranchesTags

Open more actions menu

Folders and files

NameName

Last commit message

Last commit date

Latest commit

History

11 Commits

11 Commits

bin

bin

docs/reference

docs/reference

scripts

scripts

skills/tacheles

skills/tacheles

src

src

tests

tests

.gitignore

.gitignore

CONTRIBUTING.md

CONTRIBUTING.md

LICENSE

LICENSE

README.md

README.md

bun.lock

bun.lock

index.ts

index.ts

package.json

package.json

tsconfig.json

tsconfig.json

Repository files navigation

An AI-writing linter: it catches the model's tells in your prose, on the exact line, offline.

Make AI-assisted writing sound like you wrote it, not a model. Tacheles flags the exact AI tells in your text and shows you how to cut them.

The slop is two things. Bloat: an LLM writes one token at a time, each the most probable next word, built to sound fluent and average, not short. So it pads, more words than the idea needs. Style: it writes in a register that reads as a machine, the em-dashes, the it's not X, it's Y, the same fifty words. Tacheles flags both, at the exact line, with the reason, and it runs on your machine: no AI, no API key, nothing uploaded, the same result every time.

It does not rewrite for you. It shows you what to cut.

Tacheles (תכל׳ס, tachles): the bottom line, the point. From Yiddish. The German Tacheles reden means to talk straight, no fluff.

See the difference

Before. A paragraph a model would hand you:

In today's landscape, it's not just about the tools, it's about delving into a robust tapestry of ideas that truly resonate with every reader.

Tacheles flags the tells: delve, robust, tapestry, the it's not X, it's Y opener, and the padding around them. Cut what it points at and you get:

Pick the tools that fit, then write something worth reading.

Same point, fewer words, none of the tells. It does not write the "after" for you; it points at what to cut so you can.

Why use it

What it gives you that the alternatives do not:

A cut list, not a score. AI detectors send your text to a server and hand back "87% AI". You cannot act on a number. Tacheles names the exact span, on the exact line, with the reason, the way a code linter does. You fix it and move on.

The fix, not just the flag. Every check maps to a rule from a working editor: Stephen King for English, Ильяхов and Нора Галь for Russian. The flag says what is wrong; the rule says how to fix it.

Multilingual. Ships with two full language packs (English and Russian) and one experimental (Hebrew); German and Spanish are in the next release. Adding a pack is data, not code, so it supports any language, not just English.

Tuned to you. Calibrate it to your own writing and it flags drift from your voice, not from a generic rule.

Private, free, repeatable. No model, no API key, nothing leaves your machine. Same text, same findings, every time.

Three kinds of tool get pointed at AI slop, and they do different jobs:

Tool What it does What you get back

A "% AI" detector Sends your text to a server and scores how AI it looks A number, like "87% AI". Nothing to act on.

A humanizer Uses an LLM to rewrite your draft and strip the AI patterns for you Edited text the model wrote: fast, but not yours, and different each run.

Tacheles (a linter) Flags the exact tell, on the line, with the rule to fix it A cut list. Deterministic and offline; you make the cut and keep your voice.

Most open-source options are the first two. This is the third: a cut list, multilingual, and tuned to you.

What's in the package

The linter. tacheles check (the cut list), tacheles measure (your writing's stylometry), tacheles compare-drafts (did the rewrite cut at least 10%, King's rule). Offline, deterministic, no API key.

The rewrite skill. A portable SKILL.md that runs the whole loop: check, rewrite against the language's tradition, re-check until clean. It also interviews you and builds your voice anchor. Drop it into Claude Code, Claude.ai, the Agent SDK, or any model.

Four profiles. essay-en, essay-ru, consulting-en-formal, technical-en. Copy one and tune your own.

Language packs. English (Stephen King) and Russian (Ильяхов / Нора Галь): a set of tells plus a rewrite procedure for each.

The guides. Rewrite procedures and the calibration and voice-anchor walkthroughs in docs/reference/.

Install

npx tacheles check draft.md # run once, no install npm install -g tacheles # or install it

Or clone the repo and run it with Bun: bun run bin/tacheles check draft.md.

Use

tacheles check draft.md

$ tacheles check draft.md draft.md (profile: essay-en)

HIGH line 3 s-banned-vocab delve HIGH line 3 s-banned-vocab robust HIGH line 7 r-reframe-opener It's not about the tools you pick, it's MEDIUM line 3 s-gpt-scaffolding Let's dive MEDIUM — s-em-dash-density 3 em-dashes / 45 words FAIL — 3 HIGH, 2 MEDIUM

One finding per line, on the exact line, the way a code linter reports. Add --json for machine-readable output to pipe into CI or another tool.

Exit code is 0 when clean and 1 when there is a HIGH finding, so you can gate it in CI or a commit hook. HIGH fails; MEDIUM and LOW are reported but do not fail.

There is also tacheles compare-drafts , which checks that a rewrite cut at least 10% (King's rule).

How it works

Three pieces:

Tells (the detectors) are the individual checks. Each one is a named pattern, a regex or a statistic, that flags one kind of slop. s-banned-vocab flags AI words like delve; r-uniform-polish flags overly even sentence rhythm. There are 43 active (one more is planned), grouped by type (surface, rhythm, concision) and by language (each language adds its own pack). All of them are data in src/tells/registry.json: an id, how it matches, its message, and a default severity. No tell is hard-coded in the engine.

Severity is HIGH, MEDIUM, or LOW per finding. HIGH fails the run (exit 1); MEDIUM and LOW are reported but never fail. That is the strictness knob.

Profiles decide which tells run, and at what severity, for a kind of writing. A profile is a JSON file: a list of tell ids with enabled and severity, plus optional per-tell params (thresholds, word-lists, exclusions).

A run reads the file (ignoring code blocks, inline code, and frontmatter), executes each tell the profile enables, and prints the findings with line numbers and severities. Same input, same output, every time.

What it catches

Every check is one named tell: a regex or a statistic that flags one kind of slop, paired with the rule for fixing it. We did not invent the rules. The English checks come from Stephen King's On Writing (kill the adverbs, prefer the active voice, cut the fancy word); the Russian pack from Ильяхов and Нора Галь. A few, before and after:

Tell Before After

s-banned-vocab leverage robust frameworks to foster a seamless journey use proven frameworks so onboarding stays simple

s-em-dash-density It worked — mostly — and then — well — it didn't. It worked, mostly. Then it didn't.

k-passive 40GB was exfiltrated. Mistakes were made. Attackers exfiltrated 40GB. We misconfigured the bucket.

k-adverbs significantly improved, rapidly evolving cut deploy time in half, changing weekly

r-sentence-triad It reports the result. It marks the failure. It writes the log. It reports the result and writes a log. Failures get marked inline.

ru-kantselyarit В целях обеспечения безопасности данный продукт является решением. Чтобы защитить данные, продукт их шифрует.

That is 6 of 39 active tells; the full set is data in src/tells/registry.json. Profiles decide which of them run, and how hard.

Model packs

It has separate checks for how different models write, so it catches the slop whatever you drafted with.

Claude leans on rhythm: it's not X, it's Y openers, bold one-liner aphorisms, em-dashes.

GPT leans on vocabulary and scaffolding: delve / robust, let's dive in, whether you're a..., and almost no em-dashes.

The same idea from each trips different checks:

$ tacheles check claude-draft.md claude-draft.md (profile: essay-en)

HIGH line 3 r-reframe-opener It's not about the framework you choose, it's HIGH line 7 r-bold-aphorism Good architecture isn't built. It's earned. MEDIUM — s-em-dash-density 2 em-dashes / 36 words FAIL — 2 HIGH, 1 MEDIUM

$ tacheles check gpt-draft.md gpt-draft.md (profile: essay-en)

HIGH line 3 s-banned-vocab tapestry HIGH line 3 s-banned-vocab robust MEDIUM line 3 s-gpt-scaffolding Let's dive MEDIUM line 5 s-whether-opener Whether you're a seasoned engineer or FAIL — 2 HIGH, 2 MEDIUM

These are tendencies, not laws: a GPT draft can still use em-dashes, a Claude draft can still say delve. Tacheles does not try to name the model that wrote your text. It flags the slop either way.

A full example

To make the difference concrete, here is the "before" essay from blader/humanizer's own README: an LLM draft packed with tells.

Great question! Here is an essay on this topic. I hope this helps!

AI-assisted coding serves as an enduring testament to the transformative potential of large language models, marking a pivotal moment in the evolution of software development. In today's rapidly evolving technological landscape, these groundbreaking tools—nestled at the intersection of research and practice—are reshaping how engineers ideate, iterate, and deliver, underscoring their vital role in modern workflows.

At its core, the value proposition is clear: streamlining processes, enhancing collaboration, and fostering alignment. It's not just about autocomplete; it's about unlocking creativity at scale, ensuring that organizations can remain agile while delivering seamless, intuitive, and powerful experiences to users. The tool serves as a catalyst. The assistant functions as a partner. The system stands as a foundation for innovation.

Industry observers have noted that adoption has accelerated from hobbyist experiments to enterprise-wide rollouts, from solo developers to cross-functional teams. The technology has been featured in The New York Times, Wired, and The Verge. Additionally, the ability to generate documentation, tests, and refactors showcases how AI can contribute to better outcomes, highlighting the intricate interplay between automation and human judgment.

• 💡 Speed: Code generation is significantly faster, reducing friction and empowering developers.
• 🚀 Quality: Output quality has been enhanced through improved training, contributing to higher standards.
• ✅ Adoption: Usage continues to grow, reflecting broader industry trends.

While specific details are limited based on available information, it could potentially be argued that these tools might have some positive effect. Despite challenges typical of emerging technologies—including hallucinations, bias, and accountability—the ecosystem continues to thrive. In order to fully realize this potential, teams must align with best practices.

In conclusion, the future looks bright. Exciting times lie ahead as we continue this journey toward excellence. Let me know if you'd like me to expand on any section!

Tacheles on the default essay-en profile:

$ tacheles check essay.md essay.md (profile: essay-en)

HIGH line 1 s-hedge-opener Great question HIGH line 3 s-banned-vocab transformative HIGH line 3 s-banned-vocab pivotal HIGH line 3 s-banned-vocab landscape HIGH line 3 k-adverbs rapidly HIGH line 5 s-banned-vocab seamless HIGH line 5 r-reframe-opener It's not just about autocomplete; it's HIGH line 7 s-ascii-only enterprise-wide HIGH line 7 s-ascii-only cross-functional HIGH line 7 s-banned-vocab intricate HIGH line 7 k-adverbs Additionally HIGH line 7 k-passive been featured HIGH line 9 k-adverbs significantly HIGH line 10 k-passive been enhanced HIGH line 13 k-adverbs potentially HIGH line 13 k-passive are limited HIGH line 13 k-passiv

[truncated for AI cost control]