Tooling for AI coding agents.
Portable agent skills. Each is a self-contained folder with a SKILL.md and supporting scripts that an agent can discover and run.
Install any skill with the gh skill command. Requires GitHub CLI 2.90.0 or later (gh skill is in public preview).
Preview a skill before installing it:
gh skill preview colbytimm/my-dev-tools d2-diagramAt project scope these agents all read the shared .agents/skills/ directory, so a single install covers all of them for the current repo:
# Project scope (this repo only) - installs to .agents/skills/
gh skill install colbytimm/my-dev-tools d2-diagram --agent github-copilot --scope project
# User scope (all repos, Copilot) - installs to ~/.copilot/skills/
gh skill install colbytimm/my-dev-tools d2-diagram --agent github-copilot --scope userUser-scope directories are per-agent. For a non-Copilot agent at user scope, pass its own --agent slug; running gh skill install colbytimm/my-dev-tools with no skill name opens an interactive picker that lists every supported agent.
Claude Code reads its own directories (.claude/skills/ at project scope, ~/.claude/skills/ at user scope):
# Project scope (this repo only)
gh skill install colbytimm/my-dev-tools d2-diagram --agent claude-code --scope project
# User scope (all repos)
gh skill install colbytimm/my-dev-tools d2-diagram --agent claude-code --scope userUnpinned installs resolve to the latest tagged release, falling back to the default branch when no release exists (this repo has no tagged release yet). Pin to a git tag or commit SHA to stay put:
gh skill install colbytimm/my-dev-tools d2-diagram --pin <tag-or-commit-sha>Pinned skills are skipped by gh skill update. To check for upstream changes across everything installed:
gh skill update --dry-runCombine multiple images (e.g. web screenshots) into a side-by-side, stacked, or grid layout for visual comparison. Adds optional per-panel title bars and emits a JSON manifest of panel positions for downstream annotation.
| Side-by-side with labels | Grid layout |
|---|---|
 |
 |
Draw annotations on an image (boxes, ellipses, arrows, labeled connectors, callout bubbles, text, highlights, and blur/blackout redaction) from a JSON spec. Coordinates can be absolute or panel-relative via an image-compose manifest, and a --caption bar keeps page/source labels off the content.
Used together: image-compose lays out a comparison and emits a manifest; image-annotate reads that manifest to draw on each panel with "panel": N, so coordinates measured in the original screenshots map straight through. Here two calculations are composed, then boxed and connected to show which values match and which differ.
Author software- and cloud-architecture diagrams as code with d2, render them to SVG/PNG/PDF, and embed them into markdown. Ships a lookup for the correct AWS/GCP/Azure service icons (whose hosted URLs are URL-encoded and impossible to guess), shared style/theme presets, and a renderer that degrades gracefully when raster export or the icon host is unavailable, falling back to SVG, or to remote icon references, instead of failing. Pure-stdlib scripts; requires the d2 CLI.
| Cloud architecture (with provider icons) | Software architecture |
|---|---|
 |
 |
An interactive, requirements-driven workflow for designing an Azure Cosmos DB for NoSQL data model. It captures access patterns, decides embed-vs-reference and partition keys, applies scaling patterns (hierarchical/synthetic keys, write sharding, data binning, TTL), and produces two artifacts (a requirements scratchpad and a final data model) grounded in current Microsoft Learn guidance. Ships stdlib-only helper scripts for RU/cost estimation and for linting a model against documented service limits.
Inspect a live Azure Cosmos DB for NoSQL account (read-only, via the Azure CLI) and emit an "observed model" plus optimization signals. It collects container configuration (partition keys, indexing, TTL, throughput) and Azure Monitor metrics (normalized RU consumption for hot-partition detection, 429 throttling, storage, document count), writing JSON in the same shape cosmosdb-datamodeling consumes. Requires az and az login; pure transform logic is unit-tested offline.
Used together: cosmosdb-inspect reverse-engineers a live account into an observed model and surfaces runtime signals (hot partitions, throttling, over-provisioning); cosmosdb-datamodeling lints that model against service limits, costs it, and drives a redesign. The two move you from "what does my account look like and where does it hurt" to "design and cost the fix."
Create or update README files that read like a maintainer wrote them. Encodes the structure of strong READMEs (lead with what it is, quick start early, demo with real output) plus a catalog of generated-sounding patterns to strip: marketing adjectives, AI vocabulary, emoji headers, rule-of-threes padding. Ships lint_voice.mjs, which runs markdownlint for mechanics and a set of custom markdownlint rules for voice (banned vocabulary, em dashes, summary openers, dead relative links).
Write, update, or audit project documentation past the front page: tutorials, how-to guides, references, and explanation pages, each page written as exactly one of those four types. Inventories existing docs, diffs them against the code to catch stale commands and dead links, and proposes a page plan before writing. Adds docs-specific voice checks on top of the shared rules: condescension words, future-tense steps, second person.
Used together: readme-writer owns the repo front page; docs-writer covers everything past it. Both ship the same voice lint config (plain markdownlint rules), so a repo can adopt one setup in CI for all its markdown.
MIT - see LICENSE.