Anti-slop agent skills

Two agent skills that stop generated work from looking generated — one for prose, one for interfaces.

Each is a slim SKILL.md with plain name + description front matter, plus reference files loaded on demand. Front matter carries no plugin manifest keys and no tool declarations, so both load in any agent that reads the Agent Skills format.

• human-writing/ — prevents robotic, inflated, repetitive, or formulaic prose while preserving the author's meaning, facts, voice, and audience.
• intentional-design/ — prevents generic AI interface patterns; requires a stated design direction before implementation, and separates absolute accessibility rules from overridable visual defaults.

Structure

human-writing/
  SKILL.md                          rules, workflow, exceptions, checklist
  references/
    tells-catalog.md                worked before/after rewrites, extended vocabulary
    genre-playbooks.md              per-genre shape and failure modes (11 genres)
    fabrication-policy.md           what counts as a fact, gap markers, ghostwriting
  checklists/
    review-passes.md                ordered edit / draft / review-only procedures
  templates/
    revision-report.md              for work under someone else's byline
  scripts/
    slop_scan.py                    literal tell finder, stdlib only

intentional-design/
  SKILL.md                          direction gate, two rule tiers, anti-patterns, checklist
  references/
    design-directions.md            nine worked directions with constraints
    layout-strategies.md            named alternatives to the default page shape
    typography.md                   scales, pairing, numerals, loading
    color-systems.md                role-first palettes, contrast, dark mode
    motion.md                       budgets, timing, reduced motion, performance
    accessibility.md                implementation and the manual check procedure
  templates/
    design-direction.md             the Step 1 form, plus a worked example
    tokens.css                      role-named token scaffold
  checklists/
    landing-page.md                 persuasion surfaces, including proof honesty
    app-screen.md                   dashboards, consoles, records screens
  scripts/
    ui_lint.py                      literal visual tell finder, stdlib only

Every reference pointer in a SKILL.md states a load condition, so files are read when they apply rather than always or never. Rules live in SKILL.md; references carry what a rule cannot — worked examples, tables, procedures. Nothing is stated twice.

Scripts

Both are optional, standard library only, and print findings without a score. A score invites tuning the work against the checker; these tell you where to look and say when the pattern is legitimate.

python human-writing/scripts/slop_scan.py DRAFT.md
python intentional-design/scripts/ui_lint.py src/

Both accept --json and --strict (exit 1 on findings). ui_lint.py also takes --cap to limit repeats per file.

What they cannot see: padding, register, invented detail, hierarchy, layout fit, or whether a number on a page is true. A clean run means nothing until the checklist in the matching SKILL.md is done.

slop_scan.py skips fenced blocks, quoted examples, and word lists, which is why it stays quiet on these files. It reports one finding in checklists/review-passes.md — three consecutive bullets opening with the same word. That is a deliberate parallel construction and a fair illustration of the tool's own disclaimer: a finding is a candidate, not a verdict.

Install

npx agent-antislop

Copies both skills into ./.claude/skills for the current project. No dependencies, no network access, no writes outside the directory it names.

npx agent-antislop --global              # ~/.claude/skills, available everywhere
npx agent-antislop --target codex        # ./.agent/skills, prints the AGENTS.md snippet
npx agent-antislop --target cursor       # ./.cursor/skills, prints a rule-file snippet
npx agent-antislop --dir ./tools/skills  # anywhere you like
npx agent-antislop --only human-writing  # one of them
npx agent-antislop --dry-run             # show the plan and stop

An existing copy is never overwritten without --force.

Claude Code

Nothing else to do. Start a new session; skills are selected from the description field, so they activate on matching work without being named.

Codex CLI

Codex reads AGENTS.md. The installer prints the lines to paste:

## Skills

- Before writing or editing any prose for people to read, read `.agent/skills/human-writing/SKILL.md` and follow it. It links reference files; load them when its stated conditions apply.
- Before building, restyling, or reviewing any UI, read `.agent/skills/intentional-design/SKILL.md` and follow it. Do not write UI code before its Step 1 direction is written down.

Cursor and other agents

Same pattern: install the folders, then reference them from the agent's rule file (.cursor/rules/*.mdc, .github/copilot-instructions.md) with the same trigger conditions.

Without npm

Copying the folders by hand works identically — they are plain Markdown.

git clone https://github.com/ELhadratiOth/agent-antislop
cp -r agent-antislop/human-writing agent-antislop/intentional-design ~/.claude/skills/

Licensing

MIT — see LICENSE. Fill in the copyright holder before publishing.

Both skills are original work informed by MIT-licensed research sources, credited in each SKILL.md. No rule text, taxonomy, banned-pattern list, numeric spec table, or scoring rubric was copied from those projects.