Skills that subtract: an anti-slop toolkit for coding agents

What it does

Paperthin is a catalog of ~30 agent skills, organized into four quadrants (depth, breadth, coil, mesh) by how many artifacts they touch and over how much time. Each skill encodes one piece of classic engineering discipline—re0 rewrites a drifted artifact as a clean v0 instead of patching it, readchk restates your instruction and only asks when a genuine fork survives, modelchk sizes the cheapest sufficient model tier before work starts. They install as skills for Claude Code, Codex, Cursor, Copilot, and a long list of others, and many fire on their own without being called.

The interesting bit

The whole catalog bets on subtraction. Every skill removes—noise, misreads, duplicated facts, bloat—and a pass that finds nothing to improve changes nothing. That restraint is the pitch: most agent skills add files and options forever; these are designed to never bloat your repo. The README even claims re0 was run on its own docs and left them untouched, as proof.

Key highlights

• re0 family covers the full lifecycle: plan, loop, memo, work, git, release, merge
• feynman presses a decision until you can explain it to a skeptic, or names the gap
• ssotize audits scattered facts and consolidates them into one home
• catchup / nba rebuild lost context from live state and return the single next action
• Skills are symlinked and auto-update; README translated into a dozen languages

Caveats

• The “[PROOF]” sections in the README are self-reported anecdotes, not benchmarks
• Installing requires an elevated shell for symlinking on some OSes—a minor rough edge
• It’s glue: markdown skill files plus a distribution mechanism, not code that does anything itself

Verdict

Worth a look if you use a skills-capable agent and are tired of AI-generated sprawl in your repos. If you don’t live in Claude Code or Cursor, there’s not much here for you.