claude-skills

Agent skills by @shaharsha - 13 production-grade Agent Skills that work in Claude Code, claude.ai, Codex, Cursor, and any other harness that reads the SKILL.md format.

MIT licensed. Built day-to-day; battle-tested in real projects.

---

Install

Claude Code (recommended)

/plugin marketplace add shaharsha/claude-skills
/plugin install shaharsha-skills@shaharsha-skills

This installs all 13 skills as one plugin. Or pick a subset:

/plugin install documents-and-decks@shaharsha-skills    # gdoc-sync + gslides-sync + gsheets + presentation-generator
/plugin install brand-and-visuals@shaharsha-skills      # brand-system + brand-assets + image-generation
/plugin install engineering-decisions@shaharsha-skills  # tech-design-doc
/plugin install building-agents@shaharsha-skills        # prompt-engineer + writing-project-memory
/plugin install utilities@shaharsha-skills              # namecheap-domains + office-render + viewing-videos

Manual (any other harness)

git clone https://github.com/shaharsha/claude-skills.git ~/shaharsha-skills
ln -s ~/shaharsha-skills/skills/<skill-name> ~/.claude/skills/<skill-name>
# repeat per skill, or symlink the whole skills/ dir

For Codex, Cursor, Gemini CLI, OpenClaw, etc., follow the same shape - point your harness at skills/<name>/SKILL.md.

---

Skills

Documents & decks

gdoc-sync

Push a local .md to an existing Google Doc, then fix the four things Google's converter breaks: #anchor links arriving as broken URL strings, other.md#anchor cross-doc refs, inline images at 1500pt+ wrecking layout, and missing RTL for Hebrew/Arabic. One Python script, stdlib + google-auth.

gslides-sync

Sister to gdoc-sync for .pptx -> existing Google Slides. Rewrites broken slide-anchor and cross-presentation links into native pageObjectId references, scales oversized images, applies RTL per text shape. Same service-account setup works for both.

gsheets

Comprehensive Sheets API v4 CLI - the Sheets-shaped sibling to gdoc-sync and gslides-sync. Read / write / append / clear cells; create / rename / duplicate / delete tabs; freeze rows, set column widths, merge ranges; style headers (bold + colored bg + frozen + filter as one atomic batch); apply borders, banding (zebra stripes), and conditional formatting; sort and filter; or drop to a batch-update escape hatch for charts, pivots, data-validation dropdowns, protected ranges, and gradient color scales. Same service-account auth as the other Google skills.

presentation-generator

Generate 16:9 PDF + PPTX decks where every slide is a custom AI-rendered image - not a templated layout with stock photos. Style locks globally via a reference image; composition varies per slide (full-bleed photo, infographic, architecture flowchart, big-number callout, timeline, quote card, etc.). Research -> narrative arc (SCQA / Duarte / Kawasaki) -> style lock -> parallel generation at concurrency 4 -> QA -> assemble. ~$3-5 in image-API spend for a 10-slide deck.

Brand & visuals

brand-system

Author a production-grade brand book + design system in one shot: long-form BRAND.md (20 sections), printable BRAND.html rendered to BRAND.pdf via Chrome headless, plus tokens.css (Tailwind v4 @theme + light/dark :root) and tokens.json (W3C DTCG, consumable by Style Dictionary / Tokens Studio). An anti-template interview won't finalize without one invented proper noun, three falsifiable principles, three real don'ts, and a 150-word voice sample. WCAG 2.2 AA audited at authoring time.

brand-assets

The mechanical-pixel sibling to brand-system. Five pipelines that run locally in seconds: vectorize (split-by-color-mask + potrace, not one-shot vtracer which muddies palettes), finalize-svg (snap fills to exact brand hexes, normalize viewBox), rasterize SVG -> pristine PNG, icon-pack (one SVG -> favicon + apple-touch-icon + PWA pack with iOS-opaque background and PWA-maskable safe area), color-audit (histogram opaque pixels per hex; fails on >1% drift). Bash + Python stdlib.

image-generation

Generate logos, icons, UI mockups, hero images, and product shots via OpenAI gpt-image-2 (default since Apr 2026 - took #1 in Image Arena by +242 pts within 12 hours of release) or Gemini Nano Banana 2 / Pro. Packages model-selection logic, provider-specific prompt grammars (OpenAI wants labeled segments + negatives; Gemini wants narrative paragraphs + positives only - mixing them up degrades outputs), asset templates, a transparent-background pipeline (gpt-image-2 + rembg), Hebrew/RTL guidance, and an iteration loop where Claude reads the saved image with vision and decides ship/edit/rewrite before showing the user.

Engineering decisions

tech-design-doc

Author technical design review documents - RFCs, ADRs, design docs, KEPs, partner-mode TDRs - sized correctly for the audience and the decision being made. Triages format first (1-2 page mini ADR vs 6-page standard RFC vs 10-20 page heavyweight KEP vs partner-mode for external dev partners), scaffolds from research-grounded templates, enforces load-bearing sections (BLUF summary, goals/non-goals with quantified targets, >=3 alternatives, cross-cutting checklist, decision log), inserts mandatory C4 + sequence diagrams in mermaid, and runs a static audit against best-practice anti-patterns. Pairs with gdoc-sync to push the finished doc to a live Google Doc for stakeholder review.

Building agents

prompt-engineer

Expert prompt-engineering reference for AI agents on Claude / GPT / Gemini APIs. Covers system-prompt structure, tool descriptions (the single highest-leverage quality factor), context engineering, provider differences (instruction placement, verbosity defaults, persona handling, temperature, caching), budget-model patterns, cross-provider compatibility, anti-patterns, and evaluation/judge-prompt design. Use when writing system prompts, tool descriptions, function-calling schemas, agent instructions, or when an agent keeps misfiring.

writing-project-memory

Author or audit a CLAUDE.md / AGENTS.md project-memory file (the file an AI coding agent reads at session start for project context). Encodes Anthropic's cardinal include/exclude rules, the under-200-line target, the falsifiability test, locations + loading semantics, AGENTS.md interop via @import, and Boris Cherny's compounding-engineering loop. Use when authoring a new CLAUDE.md, auditing an existing one for bloat, trimming a too-long file, or diagnosing why Claude isn't following its instructions.

Utilities

namecheap-domains

Check domain availability via Namecheap's domains.check API. One domain, batches up to 50, or TLD sweeps (com,io,ai,dev,co,app,xyz). Surfaces premium and EAP fees so you don't fall in love with a $2,999 "available" name. Stdlib-only Python; auto-chunks lists >50 (the API's hard cap).

office-render

Render a Microsoft Office file (.docx / .pptx / .xlsx) to PDF and then to page images using the real installed Office app (Word / PowerPoint / Excel) on macOS — pixel-faithful, unlike LibreOffice, which substitutes fonts and re-flows complex tables and slide grids. One command produces page JPGs Claude can read, for previewing a doc or visually QA-ing a generated .docx/.pptx. Bakes in the macOS permission gotchas it took real debugging to find: Automation consent, the Office sandbox's per-file access prompt, Full Disk Access for the Office apps (and the quit-and-relaunch needed for it to apply), and the with timeout wrapper that prevents AppleEvent timeouts on PDF export.

viewing-videos

Claude can't watch video — this skill is how it sees one anyway: turn the file into the fewest frames that answer the question via ffmpeg. Three strategies (targeted seek for known timestamps, scene-change detection for screen shares / slides, interval sampling for continuous motion), frames renamed to their meeting timestamp, an ImageMagick contact sheet to triage 60+ frames in one look, and a crop+upscale recipe for table text at the edge of 720p legibility. Born from a real need: a 30-min Zoom recording whose transcript missed the numbers shown on screen.

---

How they compose

A few of these are designed to work together:

• presentation-generator calls image-generation for every slide.
• presentation-generator consumes brand-system's BRAND.md for palette / typography / motif lock when present in the directory.
• brand-system (the document) and brand-assets (the pixels) are siblings - run both for a complete brand rollout.
• gdoc-sync, gslides-sync, and gsheets share Google service-account setup; one SA works for all three APIs.
• tech-design-doc calls gdoc-sync at the end of the workflow to push the finished TDR to a live Google Doc for stakeholder comments.

---

Compatibility

These are standard Agent Skills (SKILL.md-format). They work in any harness that reads the format:

• Claude Code (CLI) - full plugin marketplace support; install via /plugin install as above.
• claude.ai - upload skills via Settings -> Capabilities -> Skills.
• Codex CLI - point Codex at the skill folder; it reads SKILL.md natively.
• Cursor - drop the skill into your ~/.cursor/skills/ (or wherever your Cursor config expects).
• Gemini CLI - point at the skill via gemini extensions install.
• OpenClaw / other harnesses - any tool that respects SKILL.md frontmatter works.

---

License

MIT. See LICENSE.