Li Human Agent Skill

li-human

Two tools live in this folder and they both actually run. Use them. Do not

eyeball this.

python3 humanize.py draft.txt --report        # clean it, show what changed
python3 detect.py draft.txt                    # score it, five checks
python3 detect.py before.txt after.txt         # prove the delta

Both read slop.json, which is the lexicon: 100+ stock words and phrases with

plain-English replacements, 17 invisible character classes, 11 typographic

substitutions, and 11 structural tells. It is meant to be edited. If the user

has a word they always use that the lexicon strips, remove it from the file.

What gets fixed automatically

1. Invisible characters. Zero-width spaces and joiners, word joiners,

soft hyphens, byte-order marks, Unicode tag characters, non-breaking and

narrow spaces. A keyboard does not produce these. They survive copy-paste,

they are invisible in every editor, and they are the single most mechanical

thing in generated text. humanize.py deletes every one, including any

remaining Unicode format character it does not have a name for.

2. Typography. Em dash to comma, en dash to hyphen, curly quotes to

straight, ellipsis to three dots, bullet character to hyphen. The em dash pass

is the one that matters: it collapses — to , and then cleans up the

double punctuation that leaves behind.

3. The slop lexicon. delve, leverage, robust, seamless, crucial, tapestry,

testament to, moreover, "in today's fast-paced world", "let that sink in" and

the rest, each swapped for a plain word, with capitalisation preserved and

URLs left untouched.

What does NOT get fixed automatically

Structural tells get flagged, not rewritten, because changing the shape of

a sentence needs judgement:

• "It's not just X, it's Y" and "not only X but also Y"
• Rule-of-three triads
• Rhetorical one-word question lines: "The result?"
• Rocket, fire, bulb, sparkle and dart emoji
• Hashtag walls
• Reflex engagement bait: "Thoughts?", "Agree?", "Who else?"
• Uniform sentence length and uniform bullet length

That list is your job. Rewrite each flagged line by hand, keeping the meaning,

then re-run detect.py. This is the part that moves the score from REVIEW to

PASS, and it is the part a script cannot do.

The five checks

detect.py scores five signals 0-100, higher is more human:

| check | what it measures | machine looks like |

| --- | --- | --- |

| BURSTINESS | sentence-length variation | every sentence the same length |

| SPECIFICITY | numbers, names, concrete markers per 100 words | abstract nouns, no figures |

| SLOP DENSITY | lexicon hits per 100 words | stock vocabulary |

| FINGERPRINT | invisible chars, em dashes, curly quotes per 1k chars | typographically perfect |

| VOICE | contractions, person, structural tells | no contractions, staged reveals |

The verdict weights the mean at 60% and the weakest single check at 40%,

because a detector only needs one signal to fire. PASS needs an overall of 70+

with no check below 55.

Say this honestly

These are five local heuristics modelled on the signals public detectors key

on. They run entirely on the user's machine and nothing is uploaded. They are

not GPTZero, Originality, Copyleaks, Winston or Turnitin, they do not call

those APIs, and they cannot promise those verdicts. Fixing what they measure

does tend to move those numbers, because they are measuring the same

underlying things. That is the claim. Do not make a bigger one on the user's

behalf, and do not tell a user their text is undetectable.

Order of operations

• humanize.py draft.txt -o clean.txt --report
• Read the structural flags. Rewrite those lines yourself.
• detect.py draft.txt clean.txt to show the before and after.
• If the verdict is not PASS, fix the weakest check named in the output and

go again. Two rounds is normal. Five means the draft was written by

formula, and the fix is a different draft, not more passes.

• Show the user the cleaned text and the score. Never the score alone.