Shape agent capabilities into your workflow.

publishing-skills derives a public release repository from a private agent-skill development folder. It stages a copy, removes machine paths and private references, scans for leaks, adds a demo-forward README, license, manifests, changelog, and installable plugin shape, then runs verification gates. The source folder is never mutated, and the workflow stops before pushing because publication is irreversible. The result remains usable as plain SKILL.md content across providers, with Claude Code plugin support as an extra path.

cloud-sandbox-vm-skills teaches agents when and how to use Box’s cloud Linux VMs for persistent servers, isolated code execution, untrusted workloads, Docker, parallel agents, tunnels, public URLs, background workers, browser control, and ephemeral CI. The skills drive Box through its JSON CLI and target tasks that benefit from persistent storage, SSH, and real networking. Box is EU-only today and is not the best fit for GPUs, sub-500ms cold starts, very large VM concurrency, or process-level forks.

icon-generator creates project marks, app icons, and favicons by starting with bold, editable SVG concepts and escalating to image rendering only for dimensional or illustrative work. It presents two or three concepts in a contact sheet that includes tiny previews, then slices the chosen master into SVG, ICO, PNG, manifest, and HTML-head files. Standalone scripts check contrast and build the set. The full conversion path targets macOS tools such as sips and qlmanage, while the SVG-first route needs no rendering account.

macos-llm-agents is a personal macOS collection for scheduled finance news, neuroscience reading, investment-bank AI viewpoints, daily vault briefs, Apple Notes synchronization, and paper reading. launchd runs the jobs, while a small governance layer injects only the required secrets, centralizes Obsidian paths, reconciles indexes, and catches up missed slots. The repository keeps private values in local templates and requires the user to configure service keys, a vault, and macOS permissions where needed.

sw-guardrails bundles a reviewer for Speechworks copy and code plus an outreach-message writer for people who stutter, speech-language pathologists, groups, NGOs, and partners. The checks cover cure claims, fluency-as-success framing, shame, audience language, metrics, events, prompts, and embedded UI text, returning severity, rationale, and a rewrite. The project emphasizes lived experience and honest product boundaries, remains under active development, and is communication guidance rather than clinical advice.

Cloud Operating Model helps an organization describe, compare, score, and recommend centralized, decentralized, federated, or platform-engineering models. Its main deliverable layers an executive assessment over technical rationale and ends with a phased transition roadmap, including ownership shifts, exit criteria, and workforce sizing; an adaptable draw.io diagram is optional. The same workflow is available as a Claude skill and GitHub Copilot prompt file, with current vendor terminology checked against live sources when needed.

Technical-Explanatory Minimalism guides agents in making illustration-led technical diagrams, schematics, cutaways, flow figures, and interactive explanations. Interactive examples are self-contained HTML files with inline SVG and vanilla JavaScript, so they can be opened without a build step and themed with CSS variables. The rules emphasize showing mechanisms, coupling views and controls, annotating in place, and keeping data honest. Verified examples cover engines, state machines, exploded views, gears, and linked plots in light and dark modes.

governance-to-automation generates an issue-driven auto-develop.sh workflow from project governance such as SOUL.md, AGENTS.md, CLAUDE.md, and MEMORY.md. The generated loop can implement, check, run dual review, fix, refactor, re-review, commit, and open a pull request, with an optional per-task test gate. Privileged execution and automatic merging remain off unless explicit flags pass a runtime confirmation. A companion repository owns the creation and maintenance of the governance files themselves.

servo structures agent development around named sources of truth and reviewers that did not produce the work. Its skills scan the project, gather relevant context, convene experts and a dissenter, shape a spec, seal a deterministic plan, watch scope during execution, close acceptance criteria, and run a non-colluding gate that returns GO, FIX, or STOP. A manifest records where project truth lives and who owns each check. The plugin is alpha; heavier backtesting and differential plan tools remain on the roadmap.

Bellwether creates a readable daily brief from congressional STOCK Act disclosures and institutional SEC 13F filings, then adds a watchlist and a paper-trading Ledger. Users can compare copied trades with the S&P 500 through backtests, inspect momentum signals, and ask Claude to fill analyst targets or ratings for watched stocks. It uses free public data and the Ledger does not place real trades. The project is educational and explicitly not financial advice.

site-visibility-report measures how a website appears in classic search, AI citations, and direct-answer surfaces, then renders the findings as a client-facing HTML and PDF report. You choose the language, measurement depth, and page scope before running its collection and scoring workflow. Bundled templates, a standalone-file builder, and headless Chrome rendering handle delivery; the manual path needs the listed Python packages.

project-guides documents a finished app, feature, or toolkit as a matched pair: a self-contained GUIDE.html for operators and an exhaustive DEVELOPER-INDEX.md for people who need to modify the code. Separate authoring roles read the real source, verify both outputs against it, and render the guide for visual inspection; an optional AI-CONTEXT.md gives a shorter orientation. The Codex edition uses native tools and local validation without requiring the Claude CLI.

shipmate helps maintainers release JavaScript, TypeScript, or Python repositories locally, focusing on why a change matters rather than copying commit subjects into a changelog. It discovers version locations, handles repositories with multiple independent contracts, checks for drift, and pauses at a human checkpoint before tagging and creating the GitHub release. It is aimed at deliberate manual releases, not fully automated commit-driven CI.

Fashion Item Extractor breaks an outfit, runway, street-style, or lookbook photo into identifiable garments and accessories, then produces a clean product image for each item. Its default workflow writes one prompt per item, crops a source rectangle as a visual anchor, and sends the crop plus prompt to Gemini or Seedream for image-to-image generation. Users can stop after analysis or prompts; final rendering needs the chosen provider key and optional Python packages.

learn-from-project reads an AI-built or unfamiliar repository and writes a LEARNING.md that explains what it does, maps its structure, and walks through core files in execution order. It separates logic that can change a result from environment or presentation details, explains the stack for non-engineers, and compares claims with actual behavior. The guide ends with an explain-back and a short list of gaps to close; output is English unless another language is requested.

relief-plate converts a high-resolution black-and-white PNG into an STL for relief printing: dark regions become raised ink surfaces and white regions remain recessed. Oversized artwork can be divided into tiles with hidden bottom dovetails so the printed relief joins across the seam, while detail test tiles expose fragile areas before a full print. The rules target a 0.4 mm nozzle and a Bambu X2D bed, and the scripts can run without the agent skill.

Governance Encoded turns AI governance material from 12 jurisdictions and eight sectors into developer-facing directive cards, code patterns, and checklists. Its three modes map requirements to a system, compare agent frameworks such as LangGraph and CrewAI, or create handoff protocols for operations spanning jurisdictions. References carry source and currency information, but the project warns that summaries may be incomplete or outdated and are not legal advice.

mutation-killing-orchestrator reads cargo-mutants survivors, ranks them by silent-regression severity, domain criticality, blast radius, cluster size, and whether a test can distinguish the mutant, then selects up to three findings. A stronger model performs the judgment and cheaper agents write one killing test per finding before a scoped re-run verifies the result. Equivalent or low-value survivors are accepted or deferred instead of chasing a perfect score, and the workflow leaves tests staged rather than committed.