Agentic Humanizer

Agentic Humanizer

A 5-pass AI humanizer. It always runs the core rewrite workflow; Slop or Not Pro adds measured on-device AI detector checks.

• Without Slop or Not: runs the full rewrite workflow.
• With Slop or Not Pro: adds on-device AI detector scoring, readability, Text Cleanup, and cleanup stats.

Slash command: /agentic-humanizer [paste text]

Inline overrides: /agentic-humanizer language=<code> variant=<spec> dialect=us|uk grade=N level=<band> tone=casual|professional|academic length=±10|exp|trim threshold=N max=N voice=off voice-skip skip-interview [paste]

language=<code> and variant=<spec> set the target language and variant (for example language=de variant=de-AT). language= without variant= uses that language's default variant from the registry (or other:<code> for an unsupported language). variant=<tag> without language= infers the base language from the variant's BCP-47 prefix (for example variant=de-AT -> language=de); if the prefix is unsupported, keep that bare prefix as the language, use the inline variant, and warn. dialect=us|uk is a legacy English alias. grade=N is English only; level=<band> (elementary|middle|high_school|college|graduate) sets the reading level for any language. See references/multilingual.md.

What this skill does

1. Runs the built-in Claude Desktop interview (no harness detection).
2. Handles inline overrides and unsupported profile commands before any rewrite.
3. Resolves rewrite preferences from inline overrides, defaults, or the interview.
4. Optionally captures a per-run writing sample and extracts a stylometric fingerprint. Voice matching does not require Slop or Not.
5. Probes whether Slop or Not Pro is reachable via MCP.
6. Runs the 5-pass humanization workflow:
   • Core mode logs unscored iterations.
   • Slop or Not Pro runs Text Cleanup, detection, and readability checks.
7. Returns the final text, loop history, highest-impact edits, and, when Slop cleanup ran, a Text Cleanup summary.

Step 1: The interview

Claude Desktop exposes ask_user_input_v0, a single-choice prompt that takes one question per call. Run the interview only when Step 3 calls for it; the question list and the answer mapping are defined here.

The interview: four or five sequential ask_user_input_v0 calls

Issue the four required questions below in sequence; do not bundle. Add the fifth voice question only when no inline voice=off or voice-skip override is present.

Before Q1, detect the source language from the pasted text with the host LLM. If the text is under ~20 words or mixed, treat the language as ambiguous: because ask_user_input_v0 caps at four options, do not list every supported language at once. Make the first call offer the three most likely languages plus "Other (a different language)"; if the user picks "Other", they name the language in their next message (resolve it against references/multilingual.md). Only after the language resolves do you ask that language's variant. Otherwise build Q1 from references/multilingual.md: present the detected language's variants plus "Other (different language)", staying within the four-option ask_user_input_v0 cap. Populate Q2's options with the detected language's metric. The calls below show the English default. Variant sets that fit the cap: English (en-US, en-GB, Other = 3); German (de-DE, de-AT, de-CH, Other = 4); Norwegian Bokmal (nb, Other = 2); Norwegian Nynorsk (nn, Other = 2); Spanish (es-ES, es-419, Other = 3); Italian, Swedish, and Danish (single variant plus Other). Norwegian Bokmal (nb) and Norwegian Nynorsk (nn) are separate languages in the registry: Bokmal returns LIX readability while Nynorsk is unsupported for both readability and detection (tells-only). Map a chosen Norwegian option to language: "nb" or language: "nn", never to language: "nb", variant: "nn". If a language ever needs more than three named variants, keep the three most common plus "Other".

ask_user_input_v0({
  questions: [
    {
      question: "Detected English. Confirm the language and variant:",
      options: ["American English (en-US)", "British English (en-GB)", "Other (different language)"]
    }
  ]
})

ask_user_input_v0({
  questions: [
    {
      question: "What reading level should the output target?",
      options: [
        "Elementary (Grade 3-5)",
        "Middle school (Grade 6-8)",
        "High school (Grade 9-11)",
        "College or professional (Grade 12+)"
      ]
    }
  ]
})

The example above shows English grade ranges. For a non-English run, replace the parenthetical ranges with the detected language's metric from the registry band table (references/multilingual.md), for example: German "High school (Wiener about 9 to 11)", Swedish "High school (LIX about 40 to 50)", Spanish "High school (Szigriszt about 55 to 65)". Never copy the English grade ranges literally for a non-English target language.

ask_user_input_v0({
  questions: [
    {
      question: "What tone should the output use?",
      options: ["Casual", "Professional", "Academic"]
    }
  ]
})

ask_user_input_v0({
  questions: [
    {
      question: "Length policy for the rewrite?",
      options: [
        "Keep within ±10% of original",
        "Allow expansion",
        "Allow trimming"
      ]
    }
  ]
})

ask_user_input_v0({
  questions: [
    {
      question: "Mimic a writing sample of yours?",
      options: ["Yes", "No"]
    }
  ]
})

After the interview

Map the chosen labels to internal variables:

• Q1 -> language and variant: read the chosen variant (American English (en-US) -> en/en-US). Other (different language) -> prompt for the language on the next turn, resolve against references/multilingual.md, then ask its variant (warn if unsupported).
• Q2 -> reading_level: Elementary -> elementary, Middle school -> middle, High school -> high_school, College or professional -> college. For English only, also set target_grade (4, 7, 10, 13). The Desktop interview collapses College and Graduate into one option to stay within the four-option cap; for non-English the loop reads reading_level.
• Q3 -> tone: lowercase the label.
• Q4 -> length_policy: Keep within ±10% of original -> ±10, Allow expansion -> exp, Allow trimming -> trim.
• Q5 -> voice choice: Yes starts Step 4 sample capture, No skips voice matching for this call.

When Q5 is Yes:

1. If Q1 was Other (different language), first capture the language from the user's next turn, resolve it and its variant against references/multilingual.md, and finalize language and variant. Only continue to step 2 after they are resolved.
2. Say exactly: "Paste 200+ words as your next message."
3. Capture the next user turn as the voice sample and return to Step 4 for validation and fingerprint extraction.

Step 2: Profile management commands

Claude Desktop skill execution is sandboxed, so this Desktop bundle does not support local saved-profile commands. If the user asks to show, reset, or set a profile, explain that saved preferences are unavailable in Claude Desktop and offer inline overrides for the current run.

When you see one of these unsupported profile commands, respond with that limitation and stop. Do not probe Slop or run the loop.

Step 3: Resolve rewrite preferences

Preference resolution order:

1. Inline overrides (language=, variant=, dialect=, grade=, level=, tone=, length=) -> use them per key; do not interview for any key that was supplied inline. Inline language=/variant= also override detection. When language= is given without variant=, set the variant to that language's default from references/multilingual.md (the first variant listed), so the pair stays consistent (for example language=de alone resolves to variant=de-DE; an unsupported language such as language=fr resolves to variant=other:fr). When variant=<tag> is given without language=, infer the base language from the variant's BCP-47 prefix (for example variant=de-AT -> language=de); if the prefix is unsupported, keep that bare prefix as the language, keep the inline variant, and warn that the language has no curated support. If an explicit language= is also present and conflicts with the variant's base language, language= wins and the run warns. For English, when level= is set without grade=, derive target_grade from the band midpoint (elementary 4, middle 7, high_school 10, college 13, graduate 17); an explicit grade=N always wins. When the resolved language is not English, grade=N is ignored. Interview runs only for the keys NOT supplied inline.
2. skip-interview flag -> skip the interview and use defaults (High school, Professional, ±10%). Detect the source language first (per Step 1) and keep it with its default variant from references/multilingual.md; fall back to English/en-US only when detection is ambiguous or no text was pasted. For English, also set target_grade 10.
3. No inline overrides for a key -> run the interview for that key.

Run the interview using the protocol in Step 1 (which detects the source language and confirms it in Q1). Capture these rewrite settings here:

• language (a base code such as en, de, es, it, sv, da, nb, nn, or other) and variant (a BCP-47 tag or other:<spec>)
• reading_level in {elementary, middle, high_school, college, graduate}; for English also target_grade (the band midpoint 4, 7, 10, 13, 17, or any integer N from inline grade=N). When level= resolves to English without grade=, derive target_grade from the band midpoint (graduate -> 17), even though the interview itself collapses College and Graduate.
• tone in {casual, professional, academic}
• length_policy in {±10, exp, trim}

After the rewrite answers, continue to Step 4.

Step 4: Resolve voice sample

Read references/voice-fingerprint.md before running this step. Set voice_active=false by default.

Voice sample resolution order:

1. Inline voice=off or voice-skip -> skip voice matching for this call.

2. Inline voice=/path/to/file.txt -> explain that Claude Desktop cannot reliably read arbitrary local paths from the user's Mac. Ask the user to paste 200+ words if they want voice matching for this run.

3. Otherwise -> use the conditional Q5 answer already captured by the interview, or ask it now if the interview did not batch it:

   "Mimic a writing sample of yours?"

   Options: Yes, No.

If Q5 is No, skip voice matching for this call.

If Q5 is Yes, or if the user chooses to paste a sample after a voice=/path warning, say exactly:

"Paste 200+ words as your next message."

Capture the next user turn as the sample. Validate it before using it:

• Under 50 words: reject it, say the sample is too short, leave voice_active=false, and continue without changing any profile.
• 50-199 words: warn that 200+ words works better, then ask whether to continue with the shorter sample or paste a longer one.
• 200+ words: use the pasted text in memory for this run. Do not persist it to disk.

For every accepted sample, use only the first 3000 words for fingerprint extraction.

Fingerprint cache:

Keep any approved fingerprint in memory for this run only. Do not cache it to disk.

Run the extraction prompt from references/voice-fingerprint.md against the host LLM. Render the JSON fingerprint and ask:

"Looks right?"

Options: Yes, Edit, Re-extract.

• Yes: set voice_active=true with the approved in-memory fingerprint.
• Edit: let the user correct the JSON inline. Validate it against the required-field list in references/voice-fingerprint.md Required fields before using it. If the edit drops a required field, refuse to use it and offer Re-extract.
• Re-extract: ask what to change, then re-run extraction with that hint.

Run the Yes/Edit/Re-extract approval gate via ask_user_input_v0.

If extraction fails, if the sample is binary or unreadable, or if no host LLM is available for the extraction prompt, set voice_active=false, add the extraction-failure footer flag for Step 7, and continue without voice matching.

Step 5: Probe Slop or Not Pro

Set slop_mode="llm-only" and slop_backend=null by default. Probing Slop selects the enhancement path only; it never decides whether the humanizer runs.

Run a real detect_text fixture call to verify both presence AND Pro tier. Only detect_text Pro-gates, so a successful numeric result confirms Pro.

Use this fixture for the probe:

In today's digital environment, organizations often adopt new software because it promises efficiency, but the real value depends on whether people can trust it. A useful tool should explain what it does, respect the user's context, and avoid turning simple decisions into complicated workflows. Clear documentation helps teams evaluate those tradeoffs before they commit time or money.

MCP path:

Call mcp__SlopOrNot__detect_text with the fixture and include_readability: true. If the tool call succeeds and the parsed response has a numeric score or ai_probability field, set slop_mode="slop-or-not-pro" and slop_backend="mcp". Treat scores from score and ai_probability as 0-1 decimals unless the value is already greater than 1. For readability, read the Flesch-Kincaid grade from readability.scores[] where kind is fleschKincaidGradeLevel.

The probe fixture above is English, so its readability kind is always fleschKincaidGradeLevel. This call only proves Pro access; discard its readability value. Source-language readability is measured in Step 6, where the returned kind depends on the source language (see references/multilingual.md).

Claude Desktop runs skills in a sandbox that cannot reach the user's machine, so there is no Slop or Not CLI fallback here: the MCP connector is the only Pro backend. If the MCP probe is unavailable or does not return a numeric score, keep slop_mode="llm-only" and continue to Step 6. Do not skip the interview, voice matching, or rewrite loop.

Step 6: Run the loop

Use the language L resolved in Step 3: any inline language=, else the base language inferred from a variant=-only override, else the detected and confirmed language from Step 1. If L is not English, read references/multilingual.md (the registry: readability formulas, band mapping, code normalization). Read references/per-iteration-strategies.md (the per-iteration cookbook), then load the tell catalogue for L's branch.

Preserve coverage. Rewrite, don't delete: keep every fact and section the source covers and preserve the core meaning. Removing an AI tell never means dropping content. Under length_policy=trim, cut only redundancy, filler, and restated points, never a unique fact or a whole section; shorten a fact's phrasing rather than removing it. Length changes come only from the resolved length_policy, never from silent omission.

• L is en: read references/patterns.md (the 33-pattern rewrite vocabulary), references/supplemental-ai-tells.md, and references/detection-guidance.md (a false-positive guard: what not to flag and which human-writing signals to preserve). Use the full detector path; read the Flesch-Kincaid grade where kind is fleschKincaidGradeLevel. If target_grade is null or unset here (for example inherited from a non-English saved profile through a partial inline override), derive it from the resolved reading_level band midpoint (elementary 4, middle 7, high_school 10, college 13, graduate 17) before the termination check; an explicit inline grade= always wins.
• L is es, de, it, sv, da, or nb: do NOT read patterns.md (English only). Read references/supplemental-ai-tells.md and references/ai-tells/<L>.md (Norwegian Bokmal uses ai-tells/no.md). Pass the normalized language_code and never pass britishize. Read whatever score kind scores[] returns, label it by that kind, and map it to a band via the registry. The AI score is n/a (detect_text returns kind: "not_english" with a null score); call analyze_readability directly for readability.
• L is nn: do NOT read patterns.md. Read references/supplemental-ai-tells.md and ai-tells/no.md, Nynorsk section. Readability is not available; the AI score is n/a; run all iterations and select by quality. Warn the user.
• L is an unsupported language: do NOT read patterns.md. Read references/supplemental-ai-tells.md only. The AI score is n/a; do not call detect_text. Call analyze_readability with the normalized language code; if scores[] returns a known kind, log it as advisory only and never use it for termination. Run all iterations and select by quality. Warn the user.

The language branch composes with, and does not replace, the 5-iteration schedule.

When voice_active=true, Iteration 2 and Iteration 5 consume the in-memory fingerprint using the contracts in references/per-iteration-strategies.md. No other iteration uses the voice fingerprint.

Constants (overridable via inline params when Slop or Not Pro is available):

• AI_THRESHOLD = 40 (override: threshold=N)
• MAX_ITER = 5 (override: max=N)
• Grade tolerance: ±1

Slop or Not Pro setup

If slop_mode="slop-or-not-pro", run Text Cleanup on the source before Iteration 0. Store:

• source_cleaned_text
• source_cleanup_stats

Use source_cleaned_text as the Iteration 0 baseline. Score and analyze the cleaned source, not the raw source.

Core setup

If slop_mode="llm-only", use the original source as Iteration 0. Do not call detect_text, analyze_readability, or clean_text.

Cleanup stats parsing

For MCP clean_text, decode content[0].text as JSON before reading:

• cleaned_text
• removed_invisibles
• punctuation_replacements
• homoglyphs_replaced
• british_substitutions

Normalize those into this internal shape:

{
  "invisibles": 0,
  "punctuation": 0,
  "homoglyphs": 0,
  "dialect_substitutions": 0
}

Termination with Slop or Not Pro

English (L is en): AI score <= AI_THRESHOLD AND the reading-level grade test passes, or after MAX_ITER. The grade test depends on how the Graduate target was set. For elementary through college, the test is always |grade - target_grade| <= 1. For the Graduate band: when Graduate was selected as a band (via level=graduate or the interview, with target_grade derived as 17), replace the grade test with range membership grade >= 15 (the FK lower edge), so a College-level grade of 14 cannot satisfy a Graduate target; when an explicit inline grade=N was given, keep the symmetric |grade - target_grade| <= 1 tolerance so the explicit target is honored. On non-convergence, return the best iteration: lowest score that meets the grade test; if none meet the grade test, lowest score outright. Do not select an iteration that dropped a source fact or section to lower the score; the Preserve coverage rule above governs the final choice.

Supported non-English (es, de, it, sv, da, nb): no AI threshold (the AI score is n/a). Terminate when the readability score for L's formula lands in the target band (see references/multilingual.md) or after MAX_ITER. On non-convergence, return the iteration closest to the target band.

Nynorsk: tells-only, no readability and no AI score. Run all MAX_ITER iterations and select by quality.

Unsupported language: tells-only with no AI score. Run all MAX_ITER iterations and select by quality. If analyze_readability returns a known score, log it as advisory only; never terminate or select the final iteration based on that advisory score.

After selecting the final iteration, run Text Cleanup on that selected text. Store final_cleanup_stats and use the cleaned text as the final output. Then run the final scoring pass gated by L, matching the loop's per-language rule: for English, run final detect_text and analyze_readability; for supported non-English (es, de, it, sv, da, nb), run final analyze_readability only (skip detect_text, which returns not_english with a null score); for Nynorsk, skip both; for an unsupported language, run final analyze_readability only and log any known score as advisory (the AI score is n/a).

Completion in Core mode

Run all five rewrite strategies once unless the source is empty or unusable. Log AI score and readability as null for every iteration. Select the final iteration by rewrite quality: preserve meaning and the source's fact and section coverage, honor the requested reading level, tone, and length, and remove the most visible AI tells from the tell files loaded for L's branch (English: references/patterns.md plus references/supplemental-ai-tells.md; supported non-English languages: references/supplemental-ai-tells.md plus references/ai-tells/<L>.md, where Norwegian Bokmal and Nynorsk both use references/ai-tells/no.md; unsupported languages: references/supplemental-ai-tells.md alone, since no per-language tell file exists).

Mid-flight Pro-gate

If any detect_text, analyze_readability, or clean_text call returns isError: true on iteration >= 1, fall through to Core mode for the remaining iterations. See references/per-iteration-strategies.md Mid-flight Pro-gate fallback.

Step 7: Output

Render this canonical block. The example shows English; the Language line and the readability column adapt to the resolved language (see the rules below the block).

## Humanized text
<final text>

## Language
English (en-US). Readability: Flesch-Kincaid grade.

## Loop history
| Iter | AI score | Readability | Strategy |
|---|---:|---:|---|
| 0 | 92% | 11.4 (College) | baseline |
| 1 | 71% | 10.8 (High school) | pattern surgery |
| 2 | 48% | 10.4 (High school) | variant + tone |
| 3 | 27% | 9.7 (High school) | grade gap |
Converged at iter 3 (<=40% AI, grade target 9 to 11).

## Text Cleanup summary
| Stage | Invisibles | Punctuation | Homoglyphs | Dialect substitutions |
|---|---:|---:|---:|---:|
| Source cleanup | 1 | 2 | 0 | 0 |
| Final cleanup | 0 | 1 | 0 | 0 |

## Highest-impact edits
- <bullet 1>
- <bullet 2>
- <bullet 3 (optional)>

Language line. Always show the resolved language and variant, plus the readability formula's display name from references/multilingual.md, for example "German (de-DE). Readability: Wiener Sachtextformel." For Norwegian Nynorsk (nn), render readability as unavailable, for example "Norwegian Nynorsk (nn). Readability: not available." For an unsupported language, render readability as advisory when analyze_readability returned a known kind, for example "French (fr-CA). Readability: advisory Flesch-Szigriszt."; otherwise render it as not available.

Readability column. The formula is named once in the Language line (use its display name from references/multilingual.md). In the loop-history column show only the value and the band in parentheses, for example "10.6 (High school)". For unsupported-language advisory readability, show only the value and "(advisory)", since there is no target band.

AI score column (non-English). Render "n/a (detector is English-only)" on the first row and "n/a" thereafter. In Core mode, render "n/a" for every row.

Convergence line (non-English with readability). Reference the band, not the AI threshold, for example "Converged at iter 3 (readability in target band: High school)." For Nynorsk, or an unsupported language with no advisory readability score, use "Completed MAX_ITER iterations (no readability available for this language; selected by quality)." For an unsupported language with an advisory readability score, use "Completed MAX_ITER iterations (readability advisory only for this language; selected by quality)."

Show Text Cleanup summary only when real Slop or Not Text Cleanup ran. Do not show backend names in the user-facing output.

If every cleanup count is zero, replace the table with:

## Text Cleanup summary
Slop or Not found no hidden characters, punctuation artifacts, homoglyphs, or dialect substitutions to clean.

When Slop or Not Pro does not converge (English), replace the convergence line with:

Did not converge below threshold in MAX_ITER iterations. Best result shown above
(iter N at S%). Re-run with `threshold=40 max=8` for a more aggressive loop,
or `tone=casual` if professional tone is constraining the rewrite.

For non-English non-convergence, replace the convergence line with:

Did not reach the target band in MAX_ITER iterations. Closest result shown above (iter N, <formula>: X.X). Re-run with a different `level=` to widen the target.

When Slop or Not Pro is unavailable, render score and grade as n/a and add this note after the history table:

> _Ran without Slop or Not Pro. Add Slop or Not Pro for on-device AI detector scoring, readability checks, Text Cleanup, and cleanup stats: <https://slopornot.ai/download>_

For mid-flight Core-mode fallback iterations, render score and grade as n/a and add this note:

> _Iterations N-M ran without on-device scoring. Local stats are unavailable for those iterations._

If voice matching was active, add this footer note:

> _Voice matched from a pasted sample for this run._

If voice extraction failed in Step 4, add this footer note instead:

> _Voice extraction failed; ran without voice match. Paste a fresh sample on your next run to retry._

Pointer files

• references/patterns.md (the 33 AI-tells, English only)
• references/detection-guidance.md (English-only false-positive guard: what not to flag, human-writing signals to preserve)
• references/supplemental-ai-tells.md (supplemental language-agnostic AI tells)
• references/multilingual.md (the multilingual readability registry)
• references/ai-tells/<code>.md (per-language tells: es, de, it, sv, da, no)
• references/per-iteration-strategies.md (the loop cookbook)
• references/voice-fingerprint.md (voice sample extraction and loop injection contracts)
• references/slop-mcp-setup.md (install guide; surface to user when they ask for on-device AI detector scoring setup)