antislop-v3.0.0

Antislop 3.0.0

Antislop 3.0.0 makes rewrites safer, audits clearer, and installation easier across supported agents.

I've spent weeks refining antislop and I hope you enjoy using it!

Before upgrading

Version 3 removes the separate antislop-audit skill. Audit mode now lives inside the main antislop skill.

Remove any old standalone copy of antislop-audit before installing 3.0.0.

The release contains two skills:

• antislop writes, edits, reviews, and audits prose.
• generate-slop creates synthetic examples for tests and demonstrations. It activates only when explicitly requested.

Rewrite without losing the source

Antislop now inventories the source before editing. It protects facts, quantities, dates, uncertainty, attribution, citations, links, code, required terminology, and the author's voice.

The repair pipeline limits changes to identified problem areas, compares the result with the source, and rolls back when a rewrite changes protected material. Missing facts stay unresolved until the author supplies them.

Audit requests remain read-only. Asking for an audit does not silently rewrite the text.

Get clearer audit results

The rule registry has grown from 146 rules in 2.0.3 to 222 rules in 3.0.0.

Audits now separate:

• Deterministic findings confirmed by the text
• Advisory findings that need context
• Human-review findings that need judgment
• Integrity defects that do not belong in the writing score

Advisory and human-review findings carry no score penalty. The Formulaic Writing Risk Score measures writing patterns, not authorship.

Version 3 changes the scoring contract, so scores from 2.x and 3.x are not directly comparable.

Match the edit to the document

Version 3 includes general, technical, marketing, fiction, and social-linkedin profiles.

Routing adjusts structural expectations for technical articles, tickets, postmortems, release notes, messages, guides, and reference material. An optional zh-CN review checks Chinese punctuation, spacing, and translationese without changing protected technical content.

Use it across more agents

The plugin package contains both skills and manifests for ChatGPT, Claude Code, and Codex.

• ChatGPT for Work plugin import has been verified.
• Codex uses the verified native .codex-plugin/plugin.json manifest.
• Claude Code includes a native plugin manifest and self-hosted marketplace metadata, but installation remains unverified.
• Kiro gets both workspace skills and the existing Power.
• Gemini and Antigravity use the portable SKILL.md package.
• Cursor gets a project rule under .cursor/rules/.
• opencode keeps its spawnable writing and audit agent.

The old Gemini-specific extension has been removed. Gemini now uses the same portable skill format as other compatible agents.

For local automation, the included read-only MCP server can score text, prepare rewrite contracts, and validate candidate rewrites without sending text to a hosted service.

Backed by more reviewed work

Version 3 adds 18 reviewed external sources to Antislop's evidence base. Each source has a recorded revision, attribution, and adoption boundary.

Imported ideas are tested locally before becoming rules. Authorship claims, detector-evasion promises, unsupported scoring thresholds, and engagement guarantees are excluded.

See the complete source and credit record.

Downloads

Asset	Use
antislop-3.0.0.zip	Standalone Antislop skill
antislop-plugin-3.0.0.zip	Plugin package with both skills. Verified with ChatGPT and Codex. Includes an unverified Claude Code manifest.
antislop-kiro-3.0.0.zip	Kiro Power

My selfhosted git remains the development repository. GitHub carries the squashed release history used for releases and plugin imports from now on.

Full changelog: antislop-v2.0.3...antislop-v3.0.0

antislop-v2.0.3

Full Changelog: antislop-audit-v2.0.2...antislop-v2.0.3

antislop-v2.0.2

Full Changelog: antislop-audit-v2.0.1...antislop-v2.0.2

antislop-v2.0.1

Full Changelog: antislop-audit-v2.0.0...antislop-v2.0.1

antislop-v2.0.0

antislop v2.0.0

Released 2026-07-28. Ships directly from v1.7.0. The planned 1.8 compatibility release never got a tag (see docs/adr/0002-ship-2-0-without-separate-1-8-release.md); its fixes landed straight into 2.0 instead. The 1.7 score formula is not preserved. Scores from 1.7 and 2.0 are not comparable.

New: Claude Code Plugin

The repo now ships as a Claude Code plugin (.claude-plugin/plugin.json, .claude-plugin/marketplace.json). Load it straight from a local clone with claude --plugin-dir ./antislop, or install from the self-hosted marketplace:

claude plugin marketplace add https://git.drunkrhin0.au/drunkrhin0/antislop.git
claude plugin install antislop@drunkrhin0

Not on Anthropic's community marketplace yet. That submission is a separate manual step.

Rule registry

rules.json is now the single source of truth: 146 rules, each with a stable ID, severity weight, detection class, profile assignment, and overlap relationships to related rules. generate.py renders the pattern reference docs from the registry, and score.py scores against it. Before this, rules lived scattered across skill markdown with no shared schema, so a rule change meant hand-editing several files and hoping they stayed in sync.

Scoring changes

• Renamed "Slop Score" to "Formulaic Writing Risk Score," with an explicit disclaimer that it measures formulaic-writing risk and cannot prove AI authorship.
• Repeated instances of the same rule now diminish: 100%, 50%, 25% weight, capped at 3x base. One overused word no longer dominates the score the way five instances of "delve" used to.
• Scores normalize to a 500-word baseline, so a 200-word snippet and a 5,000-word report are comparable.
• Overlap handling: when multiple rules fire on the same text span, only the primary finding scores. Related findings still surface in the report but don't double-count.

Writing profiles

Two profiles ship: general (all rules active) and technical (documentation and API references, where terms like "significant" and "robust" aren't penalized). Add a profile by following _profile_guide in rules.json.

Auto-fix

fix.py / fix.sh add regenerate-and-reverify auto-fix for propagation drift: when a rule change causes derivative files (GEMINI.md, the opencode agent, pattern-reference docs) to fall out of sync, fix reruns the generator and reverifies rather than requiring a manual patch per file.

Validation

validate.py adds a repository invariant validator. Checks cross-file version parity, dash parity between original and derivative skill files, and required metadata sections. Backed by a new test suite (tests/test_fix.py, test_generate.py, test_profiles.py, test_score.py, test_validate.py) and fixture set under tests/fixtures/.

New pattern source

Added petergyang/no-ai-slop (MIT) as a credited source: padding adverbs, colon-reveal constructions, expert-cosplay phrasing, voice-preservation framing, and the protect-the-fact editing principle.

Fixes bundled into this release

• Forgejo release workflow was firing release steps on every push to main, not just tags.
• score.py silently scored clean text when given an unknown --profile instead of rejecting it. Now rejects.
• Stale docs corrected before merge: triage-labels.md described labels that don't exist on the repo, README compared 2.0 against a 1.8 release that never shipped.
• CI checkout ref corrected, tracked scratch docs dropped, artifact paths normalized.

Full Changelog: antislop-v1.3.2...antislop-v2.0.0

Full Changelog: antislop-v1.3.2...antislop-v2.0.0

antislop-v1.7.0

Full Changelog: antislop-v1.3.2...antislop-v1.7.0

antislop-v1.6.0

Full Changelog: antislop-v1.3.2...antislop-v1.6.0

antislop v1.5.0

Full Changelog: antislop-v1.3.2...antislop-v1.5.0

antislop v1.4.0

Full Changelog: antislop-v1.3.2...antislop-v1.4.0

antislop v1.3.2

Full Changelog: antislop-v1.3...antislop-v1.3.2