Skip to content
cccskills
FreePublished on GitHub

anti-slop

Rules against AI slop in code, comments, commit messages and docs. Use whenever you write or edit code, write documentation, a README, a commit message or a PR description, or review changes for quality.

The English summary is being prepared. Showing the original description.

See how to install

Files included (1)

  • SKILL.md2.9 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

Anti-slop

Slop is output that looks finished but isn't, or that pads the work with words and code nobody asked for. Hold every change to these rules.

Code

  • Do what was asked, and only that. No extra features, options, config flags, abstraction layers or "future-proofing" the task doesn't need.
  • No stubs. Never leave TODO, FIXME, pass, throw new Error("not implemented"), "implement this", "your logic here" or "... rest of the code". Write the real code, or say plainly what you didn't do and why.
  • No debug leftovers. Remove console.log, print debugging and debugger statements you added. Real logging goes through the project's logger.
  • No commented-out code. Delete it; git keeps history.
  • Comments explain why, not what. Don't narrate code ("This function adds two numbers", "Increment i"). Keep comments that record a reason, a constraint or a non-obvious decision.
  • Match the codebase. Follow the file's naming, structure, error handling and comment density. Reuse existing helpers instead of writing near-duplicates.
  • No new files without a reason. Don't add READMEs, example files, summary markdown or test scaffolding unless asked.
  • Don't swallow errors. No empty catch, no except: pass, no silent fallbacks that hide a failure.
  • Tests test behaviour. No tests that only assert a mock was called, no snapshot dumps nobody reads, no tests edited to pass without fixing the cause.
  • Finish the loop. Run the project's typecheck, lint and tests for what you touched. Report failures honestly.

Writing (docs, READMEs, commit messages, PR descriptions, replies)

  • Lead with the point. First sentence says what changed or what the answer is.
  • Plain words. Not: delve, leverage, utilize, robust, seamless, cutting-edge, game-changer, supercharge, elevate, unlock, embark, tapestry, "in today's fast-paced world", "it's worth noting", "a testament to".
  • No filler. Cut openers ("Great question!", "Certainly!"), recaps of what the reader just read, and closing offers of more help.
  • No hype or hedging stacks. Say what it does. Qualify only when it matters.
  • Formatting serves the reader. No emoji in headings, no bold on every other phrase, no bullet lists for things that read better as a sentence.
  • Commit messages: imperative subject under ~70 characters, then what and why. No "This commit...".
  • Be specific. Name the file, the function, the number. "Improved performance" says nothing; "cut cold start from 2.1 s to 0.8 s" does.

Before you finish

Re-read your diff as a reviewer would. Remove anything on the lists above. If the ADE plugin's slop check reports a finding, fix it unless it's intended, and say so in one line when it is.

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

AI Agent Architect: Design agent systems: single vs multi-agent, tools and MCP, memory, guardrails, evals.

The English summary is being prepared. Showing the original description.

alvin-reyes/better-agentic-ide85Updated Oct 18, 2026

AI Automation Strategist: Decide what to automate: opportunity mapping, ROI, build vs buy, rollout and change management.

The English summary is being prepared. Showing the original description.

alvin-reyes/better-agentic-ide85Updated Oct 18, 2026

AI x Web3 Architect: Autonomous agents that hold wallets: smart accounts, spend limits, intents, agent payments.

The English summary is being prepared. Showing the original description.

alvin-reyes/better-agentic-ide85Updated Oct 18, 2026

Workflow Automation Architect: Automate business processes: triggers, queues, retries, approvals, n8n vs Temporal vs code.

The English summary is being prepared. Showing the original description.

alvin-reyes/better-agentic-ide85Updated Oct 18, 2026

Smart Contract Systems Architect: Contract system design: modules, upgradeability, access control, storage, audit readiness.

The English summary is being prepared. Showing the original description.

alvin-reyes/better-agentic-ide85Updated Oct 18, 2026

Cross-chain & L2 Architect: Chain selection, rollups, bridges and cross-chain messaging with their trust assumptions.

The English summary is being prepared. Showing the original description.

alvin-reyes/better-agentic-ide85Updated Oct 18, 2026

See all skills by alvin-reyes

Report a problem with this skill