Paperthin bets that the best agentic skill is knowing when to subtract

Paperthin starts from a diagnosis that most agent skill libraries get backwards: agents are rewarded for adding — more files, more options, more boilerplate — and nothing in the default loop ever makes them subtract. Every skill here bets the other way. The collection is organized around a two-axis map (one artifact vs. many; now vs. across iterations) and the skills follow that structure into four quadrants: depth, breadth, coil, and mesh.

The founding skill is re0, which rewrites a drifted artifact as a clean v0 rather than patching it again. That single instinct — rewrite, don't accumulate — runs through the whole catalog. debloat compresses to load-bearing density. detool strips incidental vendor nouns from portable content, leaving the mechanism they meant. dedash removes the em-dash tell one judged occurrence at a time. shower hands the artifact to a context-free sub-session and asks whether it stands alone — and when tested on its own spec, it caught three bugs the author had missed.

The epistemically interesting skills are the ones that attack self-confirmation. hate returns the single load-bearing objection that could kill a plan plus the cheapest test that would prove it matters — nothing else. mandela audits evaluations against an eight-pattern leakage taxonomy, because a model, a scorer, and a designer can all agree on a result while no outside ground-truth ever entered. macrothink fans out two to five independent reads and reports divergence first, explicitly banning majority vote and "verified" wording from same-model convergence.

The coil quadrant handles what happens across iterations. re0-loop runs a build → QA → re0-memo → re0-work cycle so lessons compound rather than code. catchup rebuilds the owner's mental model from live state when they return after a long autonomous run — coined terms, renamed files, and all. nba then returns the single next best action. The README is explicit that a correct next move is not a briefing: the two skills must run in order.

Installation is a single npx command that symlinks skills globally across whatever agents you use, with re0-upgrade handling catalog updates. The restraint claim is tested directly: the README documents running re0 on its own docs when they were already at v0, and it changed nothing. A tool that finds nothing to improve when nothing is wrong is the stated goal.