v1.9.0

Changed

• RHET-8 now covers landing lines at sentence, paragraph, and section ends:
  clever buttons, performed insights, quotable closers, heroic stamps, and
  slogan resolutions. Its repair is to delete the flourish or replace it with
  a checkable detail already supported by the source.
• SUBST-1 now requires a specific, checkable detail for every claim.
  SUBST-4 names the mechanism, failure mode, reason, cost, and trade-off as
  the facts that make a decision useful to the reader. The fix contract now
  cuts or flags claims whose supporting detail is absent from the source; it
  never invents the missing fact.
• STRUCT-3 no longer tells writers to end short for effect, and REVISE-9
  no longer treats a closing line as something that automatically deserves
  polishing. Both now defer to RHET-8.
• TONE-3 now states the builder-to-peer test without using the contrast pivot
  that RHET-1 bans.
• The final-pass checklist now checks landing lines, justification tails,
  exact-background claims, watchlist vocabulary, announced honesty, and the
  stronger specificity and name-swap tests directly.
• The runtime manifest and skill metadata now report version 1.9.0.

Source note

The trigger was a set of repeated corrections from working prose. The strongest
recurring tell was the sentence trying to finish on a polished button: "happiest
a step behind the thing everyone showed up for," "so I know what that bar costs
to hit," and "one typed line." The same review confirmed that the surrounding
specificity, cadence, vocabulary, and voice rules already belonged to the
catalog and needed sharper application. No duplicate IDs were added.
A blind forward test caught one more gap before release: the exact sentence
"We improved performance" escaped because the old SUBST-1 example included
the extra word "significantly." The canonical example now matches the shorter
form.

v1.8.0

Added

• A Codex skill-only plugin manifest at .codex-plugin/plugin.json, with
  complete publisher, discovery, capability, and starter-prompt metadata.
• A repository marketplace at .agents/plugins/marketplace.json, so Codex can
  install Slopspeare directly from davidvictor/slopspeare.
• Explicit per-agent install examples and a public skills.sh catalog badge.
• Contract checks that keep the Agent Skill, Codex plugin, marketplace, and
  changelog versions and identifiers aligned.

Changed

• The runtime manifest and skill metadata now report version 1.8.0.
• Installation guidance now distinguishes the portable skills CLI route from
  the Codex-native plugin route and names the update behavior of each.

Source note

The trigger was a distribution audit against the open Agent Skills
specification, OpenAI's current skill-only plugin guidance, the skills CLI,
and public repositories from OpenAI, Anthropic, Vercel, and NVIDIA. The
portable skills/slopspeare/ package remains canonical; the Codex plugin is a
thin native wrapper over that same directory.

v1.7.0

Added

• RHET-17 (cut participial analysis tails that merely announce
  significance). This catches trailing "highlighting," "underscoring,"
  "reflecting," and "showcasing" clauses that attach praise to a fact without
  explaining a mechanism or result.
• SUBST-18 (name the source behind attributed claims). "Experts agree,"
  "studies show," and similar borrowed authority now require a verifiable
  source. A missing source is cut or flagged; it is never invented.
• STRUCT-10 (make verbs carry the action). Nominalizations and weak phrases
  such as "made a decision" and "has the ability to" now resolve to direct
  verbs when meaning is preserved.
• EDIT-EVAL.md, a post-edit fidelity check covering meaning, evidence, voice,
  proportionality, structure, audience fit, and reporting integrity.
• Contract tests for the three new rules, expanded vocabulary, edit fidelity,
  authorship boundaries, and public package contents.

Changed

• The fix workflow now reads the full draft, infers or obtains audience and
  delivery context, creates an internal voice fingerprint, makes the smallest
  effective changes, preserves useful roughness and structure, reports
  meaningful reorganization, and passes EDIT-EVAL.md before delivery.
• Scan, comment, and score outputs now treat named patterns as observable
  evidence. They never claim or estimate whether AI authored the source.
  score is explicitly a catalog coverage and finding-count receipt.
• VOCAB-2, VOCAB-3, VOCAB-4, VOCAB-5, VOCAB-8, and TONE-14 gained
  the missing vocabulary, filler phrases, contextual adverbs, and honesty
  variants found during a comparison with
  petergyang/no-ai-slop.
• FORMAT-5 now covers headers over tiny sections and bullet lists that should
  be continuous prose.
• The runtime manifest and skill metadata now report version 1.7.0. Generated
  package metadata includes EDIT-EVAL.md.

Source note

The trigger was a rule-by-rule comparison with Peter Yang's MIT-licensed
no-ai-slop skill. Slopspeare retained its existing severity model, audit
coverage contract, and stricter punctuation decisions while adopting the
transferable editorial safeguards named above.

v1.6.0

Changed

• Bare invocation and bare scan now run a full deep audit of every active rule,
  including guidelines. scan all remains an explicit alias for the default.
• Audits can be scoped by category, strictness, exact rule IDs, or intersections
  of those selectors. The abbreviated former default is now the explicit quick
  command.
• Full and scoped audits now build a per-rule coverage ledger and report checked
  rule and category counts, so a result cannot claim full coverage after sampling
  only the checklist or quick reference.
• The runtime loads rules.json as the active audit manifest and opens canonical
  RULES.md entries only when their structured form needs clarification.
• Skill metadata now follows the portable Agent Skills shape, with the version
  under metadata and no implementation-specific tool allowlist.

Added

• Codex-facing agents/openai.yaml metadata.
• Contract tests for the full-audit default, selector grammar, coverage protocol,
  portable frontmatter, and generated public package.

v1.1.0

Installable skill

Slopspeare is now an installable agent skill.

npx skills add davidvictor/slopspeare

It works as a linter. It reads prose, cites each AI tell by rule ID and severity, and offers a minimal fix. It can apply the fixes too.

Commands

• scan (default). Audit a draft and report findings by rule ID and severity. scan all includes the guideline tier.
• fix. Edit the draft. Bare fix applies hard bans and strong defaults and reports a per-rule changelog. fix hard, fix all, and fix RHET-1,PUNCT-1 set the scope. Given a file path, it edits in place.
• score. The tier counts and a one-line read.
• comment. Findings as paste-ready review citations.

House style

Drop a slopspeare.local.md in your repo to layer project-specific rules on top of the public catalog. See slopspeare.local.example.md.

The 81-rule catalog is unchanged from v1.0.0.

v1.0.0

Initial public release.

Slopspeare is a set of rules for spotting the patterns that make writing read as machine-generated, and for replacing them with writing that reads as a specific human.

• The rules, each with the reason it exists and a before/after
• A printable final-pass checklist with grep strings
• A one-screen quick-reference cheat sheet
• A conflict log for places where sources disagreed
• A machine-readable rules.json for tooling

Licensed CC BY 4.0.