Shape agent capabilities into your workflow.

REM Sleep processes an agent’s session trace into atomic, retrievable memories, links related lessons, prunes noise, and separates useful guidance from the emotional charge of an event. Its default store is a file-based index with one fact per file and frontmatter, but the stages can adapt to another format after asking once. The skill is intended for cross-session learning and works with Claude Code, Codex, and Cursor without requiring a database.

Immune Check scans an artifact before it enters a public repository, external service, release, or inbox. It looks for API keys, private keys, passwords, cookies, personal or health data, internal hostnames, home paths, build output, and oversized vendored material, selecting an available scanner or falling back to a built-in pattern pass. The skill is stack-agnostic and records which scanner ran, so a release gate can show the depth of its check.

Sunset guides an agent through deciding whether a project is truly finished, recording what worked and what ended it, harvesting reusable code or lessons, and archiving the project with a visible marker instead of deleting it silently. It also closes lingering loops so an inactive project stops consuming attention. The archive is designed to be reversible, and the first safeguard is distinguishing a frozen project from one that is genuinely dead.

Roam turns a natural-language world brief into a single HTML experience that can be explored in the browser. One track builds dependency-free CSS-3D corridors for scroll-through stories; another uses Three.js for first-person rooms with WASD, mouse look, GLB furniture, and tileable textures loaded through an import map. The skill avoids a build step, but the project is an early prototype and the room track still depends on browser-loaded Three.js assets.

polish packages several agent skills for Claude Code, Codex CLI, and Cursor; its prompt-polish entry turns a rough question into a ready-to-paste bilingual prompt by selecting suitable prompt-engineering strategies. The same collection includes duplicate-file planning, a pre-push gate, frontend design, brainstorming, test-driven development, and systematic debugging. Users should invoke the relevant skill by name, because the repository is a collection rather than one monolithic prompt editor.

Playtime gives an agent a structured way to explore before a capability is needed. It chooses a playground with no production blast radius, tries unfamiliar tools, builds throwaway experiments, challenges skills with strange inputs, and harvests useful surprises into durable learning. The practice complements memory consolidation by generating new experience on purpose, while every experiment stays inside the system’s boundary. It is developmental work, not a request to ship the experiment.

Contemplation asks an agent to hold a question open instead of rushing toward an answer. It checks whether the frame is right, examines the goal and values behind it, names both sides of a tension at full strength, and can conclude that the tension should remain unresolved. The practice is solo and non-adversarial, distinct from a debate council, task retrospective, or memory pass. It is structured deliberation, not a mystical claim.

certora-prover-skill helps Claude Code write Certora Verification Language specifications for Solidity and EVM contracts, including rules, invariants, ghosts, hooks, and method summaries, then run certoraRun and diagnose counterexamples, vacuity, soundness, and timeouts. It keeps the distinction from testing explicit: a result covers every state allowed by the model, not every behavior omitted from the specification. The skill is an unofficial community package and requires the Certora Prover environment.

assay interviews the user until scope, repositories, contracts, and testable criteria are explicit, then runs an autonomous plan, research, implementation, convergence, polish, and ship-gate workflow. Machine checks decide whether correctness and completeness are sufficient, while the human intervenes at intake and before outward actions such as commits, pull requests, or ticket writes. The skill adapts to backend, frontend, mobile, CLI, library, and infrastructure projects across one or several repositories.

Solution Ladder makes an agent consider several architectural responses to a refactor or wiring problem: a local patch, a per-consumer escape hatch, a safer right default, and an outside-the-ladder wildcard. Separate blind subagents develop the options without knowing one another, then a blind judge ranks them against the current codebase and flags both insufficiency and overreach. The skill works in Claude Code, Codex CLI, and OpenCode; no level is universally correct.

sign-safe decodes an opaque Solana transaction before signing, classifies dangerous instruction shapes from the signer's perspective, and emits a SIGN, HOLD, or REJECT verdict plus machine-readable JSON. Its core requires no key, RPC, or broadcast; optional simulation and RPC enrichment are separate, and SIGN only means the transaction passed the current static policy.

Trim Diff Noise classifies each branch diff hunk as definite noise, necessary, or borderline before changing anything. It reverts only approved definite noise, asks about borderline cases, verifies each revert with the project's test and lint commands, and leaves all reverts unstaged for review; its default verification assumes Ruby with RSpec and RuboCop.

apple-hig reviews interfaces for iOS, iPadOS, macOS, watchOS, tvOS, and visionOS against Apple Human Interface Guidelines. It looks at platform-specific interaction, accessibility, design tokens, and SwiftUI code, then reports actionable issues; the project lists Claude Code and an internet connection for current guideline data as requirements.

zx-skills bundles modular skills for Sinclair BASIC, C, Z80 assembly, graphics, AY sound, special display modes, and ZEsarUX control. They support a build-load-run-debug workflow for ZX Spectrum and Timex projects, while actual compilation or emulator control still depends on the relevant toolchains, emulator, and optional MCP server.

nrf-proposal-skill organizes Korean NRF research proposals around program scoring, literature research, research gaps, logic structure, Korean drafting, figures, panel simulation, and HWPX export. It marks unverifiable claims, but the project explicitly treats output as a learning or reference draft that applicants must rewrite and verify rather than submit unchanged.

linkedin-writer calibrates a user's voice, researches recent industry updates, interviews them for firsthand material, proposes a weekly plan, and writes approved posts with image descriptions. It does not invent numbers, situations, or personal details, and stores the voice profile in VOICE.md for later sessions; a Claude Code account is required.

product-growth-skills first diagnoses where a product should focus by building a driver tree and ranking leverage, influence, and compounding. Its experiment-design plugin then proposes candidates, scores them with ICE, and shapes a top-three portfolio with hypotheses, metrics, guardrails, sample or duration guidance, and right-sized statistics; it does not run experiments or analyze their results.

paper-maestro is a Claude Code orchestrator for moving an academic manuscript from idea and literature gap through structure, drafting, data, citation verification, review, and submission. It is tuned to Russian VAK and GOST requirements, with a separate English or Q1 track, multi-agent barriers, and quality gates for overclaiming, reproducibility, and unsupported citations.