LilMGenius/paperthin

Low-level agentic design patterns. Turning old engineering wisdom into reflexes your agent reaches for on its own—on any agent

Summary

Paperthin is a JavaScript repository of "skills" — discrete design patterns that shape how coding agents draft, edit, and review artifacts. Each skill is framed around removing noise rather than adding features, and they are organized into four quadrants: single-artifact depth, cross-document breadth, iterative coil, and multi-perspective mesh. The repository is released under the MIT license and is not archived.

What it is useful for

The project ships 28 skills, each targeting a specific failure mode in agent-driven work:

• Single-artifact cleanup (depth/) — re0 rewrites a drifted document into a clean v0 instead of patching it; debloat compresses prose to load-bearing density; dedash strips em-dashes; shower cold-reads with zero context; factchk verifies claims against external sources in both directions; modelchk recommends a capability tier (fast/standard/frontier) and reasoning effort.
• Cross-document consistency (breadth/) — ssotize finds scattered copies of one fact and consolidates them under one canonical home.
• Iterations that learn (coil/) — re0-plan, re0-loop, re0-memo, and re0-work structure a build → QA → memo → restart cycle, with nba returning the single next best action from live state.
• Multi-perspective review (mesh/) — prism runs independent lenses and reports only where they clash.

The README lists supported agents as Claude Code, Codex, OpenCode, Antigravity, Copilot, Cursor, Grok-Build, Pi, Hermes, and OpenClaw.

How engineers can use it

The documented install runs through a separate CLI:

npx skills@latest add LilMGenius/paperthin --global --agent '*'

The README specifies running this from an elevated or admin shell so the skills are symlinked (auto-updating) rather than copied. Invoke a skill by name (e.g., /re0, /debloat, /factchk) or let model-invoked skills fire on their own; /re0-upgrade refreshes the catalog.

Documented caveats and limitations:

• The install depends on the third-party skills package, not on Paperthin directly. The README does not document an alternative install path if that tool is unavailable.
• Skills marked Read-only ✔ in the Index report but do not mutate. Invocation matters: user-invoked skills run only when called, while model-invoked ones may fire automatically.
• The project's stated success criterion is restraint — a skill that finds nothing to change is treated as working correctly, so deliberate no-ops are expected.
• The README ships in 11 languages; the English version is the canonical source for skill definitions.