Slop Check

Slop Check

One-shot access to Slop or Not Pro's on-device analysis. Detect AI text or images, clean AI artifacts, or score readability, then report a clear verdict. No interview, no loop.

Slash command: /slop-check [paste text | file path | image path]

Critical rules

1. For image-detect and explicit image-score, MUST prefer the app-bundle CLI when the current client can run local shell commands. Use MCP for image operations only when CLI execution is unsupported or fails. For text, readability, cleanup, and status, try MCP before CLI.
2. MUST verify Pro with a real Pro-gated call. NEVER infer Pro from slop_status or slop status; both succeed for non-Pro users.
3. NEVER block on a structured-question tool. Resolve ambiguity with the Step 1 precedence rule and state the assumption, or emit the single plain-text input request. Plain text renders on every harness.
4. NEVER create symlinks, edit shell rc files, or otherwise modify PATH. Prefer the absolute app-bundle binary at /Applications/Slop Or Not.app/Contents/MacOS/slop whenever using the CLI. Detection, readability, cleanup, score, and status are output-only: run them immediately, no permission needed.
5. Detection scores are 0-1 decimals; MUST multiply them by 100 for display. Readability values are grade/ease numbers and MUST NOT be converted to percentages. NEVER guess a flag or JSON field path; look it up in references/slop-tools.md.
6. Image checks MUST default to detect_image. Use score_image only when the user explicitly asks for a raw or absolute OmniAID score. A request to "score this image" without OmniAID still maps to image-detect.

Step 1: Identify the operation and input

Pick one operation from the request. Do not ask a question to disambiguate.

Resolve the input by precedence (first match wins):

1. A path ending .png, .jpg, .jpeg, .heic, or .webp, or an explicit image reference: image operation on that file.
2. Any other explicit file path: read that file as text.
3. Inline pasted text in the request: use it.
4. The user says to use the clipboard: pbpaste.

Ambiguity rule (NEVER ask via a tool): if no operation verb is present, default to text-detect for text input and image-detect for image input, and state the assumption in the output, for example: "Assumed AI-detection. Ask for readability or cleanup to switch."

Only if there is no usable input at all, emit exactly this one line, then stop and wait:

Paste the text, or give a file or image path, and say detect / clean / readability.

Step 2: Resolve a working backend

Use this backend order. Stop at the first that works.

1. Image operations with shell support: CLI first. For image-detect and image-score, first check the app-bundle binary:

   ls "/Applications/Slop Or Not.app/Contents/MacOS/slop"
   

   If it exists and the client can run shell commands, use that quoted binary path. This avoids converting local image files to base64. If the client has no shell or CLI execution, for example MCP-only clients such as Claude Desktop or ChatGPT connectors, skip to MCP.

2. MCP for non-image operations, or image fallback. Call the operation's SlopOrNot MCP tool (Claude Code surfaces these as mcp__SlopOrNot__<tool>; other harnesses expose the same SlopOrNot server tools). If it returns a result and is not a Pro-required error (isError: true or a Pro-required message), use it. For status, slop_status is only a health/version call: also run the Pro proof probe in references/slop-tools.md section 5 before reporting active Pro. Done.

3. CLI fallback for non-image operations. If MCP is absent or errored, check whether the app-bundle binary exists:

   ls "/Applications/Slop Or Not.app/Contents/MacOS/slop"
   

   • If it exists: complete this turn's operation by calling the binary at its absolute quoted path. Do not modify PATH.
   • If it does not exist: this is genuinely not installed. Go to the Step 5 "Not installed" fallback. For status, run the absolute binary's status --json plus the Pro proof probe before reporting active Pro.

4. Pro-gated failure. If the CLI runs but a Pro-gated call exits non-zero or returns Pro-required, the app is installed and Pro is not active: go to the Step 5 "Installed but Pro is not active" fallback. For status, report Pro as inactive per Step 4 instead.

Tool names, flags, parameters, and JSON field paths are in references/slop-tools.md. Setup and registration details are in references/slop-setup.md. Consult them; do not guess.

Step 3: Run the operation

Use the operation-to-tool map in references/slop-tools.md section 5. Key points:

• text-detect: request readability alongside detection (include_readability: true on MCP; slop text --json returns both). When the input language is detectable, pass a normalized language_code (Norwegian Bokmal is nb; never pass the macro no, remap it to nb); omit it when ambiguous so the app auto-detects. Passing nn (Nynorsk) is harmless but returns kind: "not_english" with a null score, same as any non-English language. For non-English input detect_text returns kind: "not_english" with a null score; see Step 4.
• readability: pass the same normalized language_code when detectable. Norwegian Bokmal is nb; always remap a detected no to nb before passing (never pass no; it returns unsupported_language:no even for Bokmal text). The returned scores[] kind depends on the language and is not always Flesch-Kincaid; read whatever kind is present (see Step 4 and references/slop-tools.md section 2).
• Images: default to image-detect. When shell commands are available, use the absolute CLI path with shell redirection from the image path: "/Applications/Slop Or Not.app/Contents/MacOS/slop" image --json < "<path>". MCP uses detect_image with base64 (base64 -i <path>) only when CLI is unsupported or fails. Use score-image / score_image only for image-score, and only when the user explicitly asks for a raw or absolute OmniAID score. If the user asks for a generic image score without OmniAID, keep image-detect and format the detection probability.
• cleanup: keep the default removals on unless the user asked to limit them. Normalize MCP and CLI cleanup fields before formatting: MCP returns cleaned_text and integer counts; CLI returns cleanedText plus count arrays.
• status: run the health call and then the Pro proof probe. Report active only when the Pro-gated probe succeeds.

Step 4: Present the result

Read score normalization in references/slop-tools.md section 2 first (detection score is a 0-1 decimal, multiply by 100; MCP verdict is a plain string, CLI verdict is the single-key object at detection.result._1).

Render the matching block.

Detection (text or image):

**Likely AI** (most_likely_ai_slop) · 87% AI probability
Language: en · Sentences: 6

Map the verdict by rule (verdicts have gradations like probably_ai_slop): real -> "Likely human"; contains ai_slop -> "Likely AI"; else title-cased. Show the label, the raw verdict in parentheses, and the percentage.

When MCP detect_text returns kind: "not_english" (non-English input), score and verdict are null. For CLI slop text --json, normalize the same case before reading detection.result: if detection.notEnglish, detection.not_english, or a normalized kind: "not_english" is present, or if the response has a non-English detectedLanguage / readability.language and no detection.result or detection.resultFewSentences, treat the AI score and verdict as null. Do not show an AI percentage. Print instead: "AI text detection is English-only; no score available for ." (read the language from the language, detectedLanguage, or readability.language field), then show the readability block if the response carries one. Never invent a score for non-English input.

Readability:

Read whatever kind scores[] returns; do not assume fleschKincaidGradeLevel. When the language returns two kinds (English: fleschKincaidGradeLevel plus fleschReadingEase), band and lead with the Flesch-Kincaid grade and show Reading Ease as a supplemental value in parentheses; never derive the band from Reading Ease, whose 0-to-100 scale is inverted. For a single-kind language, band that one kind. Look up the formula display name, scale direction, and band ranges in references/slop-tools.md section 2. Show the formula name, the value, the detected language, and the band.

**Flesch-Kincaid grade: 9.7** (Reading Ease 62.4)
Language: en · Band: High school · Words: 210 · Sentences: 14

For non-English, label by the returned kind, for example:

**Wiener Sachtextformel: 11.2**
Language: de · Band: High school · Words: 180 · Sentences: 12

If scores[] is empty, read warnings. Warnings come back in two shapes by backend: MCP returns colon-tagged strings (unsupported_language:<code>, insufficient_text:NN, approximate_syllable_counts:<lang>); the CLI returns camelCase objects under readability.warnings[] (unsupportedLanguage, insufficientText, optionally with wordCount). Treat the two forms as equivalent: unsupported_language / unsupportedLanguage means no score is available (report "Readability not available for in this app version"); insufficient_text / insufficientText is a soft warning (scores may still be present; treat as advisory), reported as "Not enough text to score reliably (insufficient_text:28)."; approximate_syllable_counts / approximateSyllableCounts is benign. Do not invent a value.

Cleanup: print the counts line, then the cleaned text in its own fenced text block so the user can copy it verbatim. For CLI output, use cleanedText as the cleaned text, sum invisibleCounts[].count, punctuationCounts[].count, and homoglyphCounts[].count, and count britishMappings.length.

**Cleaned.** Invisibles: 3 · Punctuation: 5 · Homoglyphs: 1 · British: 0, followed by a fenced text block containing only cleaned_text.

Image score (only for explicit OmniAID requests):

**Raw image score: 0.80** (OmniAID)

Status:

**Slop or Not Pro: active** (version 1.0.9)

Use inactive when the health call works but the Pro proof probe returns a Pro-required error.

When the ambiguity rule fired, prepend the one-line assumption note.

Step 5: Fallback when Slop or Not Pro is unavailable

Two distinct cases. Pick the message that matches the real state. Never tell an installed user to install the app.

Not installed (MCP is down AND the app bundle is absent). End with exactly this paragraph:

Slop or Not is not installed. Install Slop or Not for Mac, then unlock Pro from inside the app: https://slopornot.ai/download

Installed but Pro is not active (the slop binary or app bundle exists, but a Pro-gated call returned Pro-required or exited non-zero). End with exactly this paragraph:

Slop or Not is installed but Pro is not active. Unlock Pro from inside the app under Settings then Subscription: https://slopornot.ai/download

In both cases, then say: see references/slop-setup.md for install, Pro unlock, and MCP/CLI registration. A missing PATH alone is handled by the app-bundle fallback in Step 2, not deflected here. For the status operation, report Slop or Not Pro: inactive per Step 4 instead of these paragraphs.

Examples

Example A, text detection (no verb, ambiguity rule fires):

User: /slop-check In today's fast-paced digital landscape, leveraging synergies is paramount.

Action: no verb, text input, default text-detect. MCP detect_text with include_readability: true.

Output:

Assumed AI-detection. Ask for readability or cleanup to switch.

**Likely AI** (most_likely_ai_slop) · 91% AI probability
Language: en · Sentences: 1

Example B, image detection from a path:

User: is this AI? ~/Desktop/art.png

Action: path ends .png, so image-detect. Shell commands are available, so use the app-bundle CLI: "/Applications/Slop Or Not.app/Contents/MacOS/slop" image --json < "$HOME/Desktop/art.png".

Output:

**Likely AI** (most_likely_ai_slop) · 80% AI probability

Example C, readability of a file:

User: what grade level is draft.md

Action: explicit file path, readability. MCP analyze_readability with the file contents and a normalized language_code when detectable. Read whatever kind scores[] returns (here English fleschKincaidGradeLevel).

Output:

**Flesch-Kincaid grade: 7.2** (Reading Ease 71.8)
Language: en · Band: Middle school · Words: 148 · Sentences: 11

For a German file, the same operation reads wienerSachtextformel4 and labels it "Wiener Sachtextformel" with the German band.

Example D, cleanup with a zero-width character:

User: clean the invisible characters out of this: Hello​world

Action: cleanup. MCP clean_text (defaults on).

Output: the counts line **Cleaned.** Invisibles: 1 · Punctuation: 0 · Homoglyphs: 0 · British: 0, then a fenced text block whose only content is Helloworld (the zero-width character removed).

Example E, explicit OmniAID score:

User: give me the raw OmniAID score for ~/Desktop/art.png

Action: explicit OmniAID request, so image-score. Shell commands are available, so use the app-bundle CLI: "/Applications/Slop Or Not.app/Contents/MacOS/slop" score-image --json < "$HOME/Desktop/art.png". If CLI execution is unsupported, use MCP score_image with base64 image input.

Output:

**Raw image score: 0.80** (OmniAID)

Pointer files

• references/slop-tools.md: full CLI and MCP surface, parameters, flags, JSON field paths, score normalization, Pro-gating, operation map.
• references/slop-setup.md: install, Pro unlock, app-bundle fallback, MCP/CLI registration, troubleshooting.