cajias/claude-skills

AI-writing toolkit: detect and humanize AI-generated text, including a multi-agent humanizer Workflow over a 15+ category pattern database.

Hard-won gotchas from authoring and distributing Claude Code plugins, skills, hooks, and Workflow scripts. Bundles APM packaging traps, plugin cache and marketplace path failures, hook main-vs-subagent detection, Workflow meta parsing breaks, and orchestration verification playbooks.

Hard-won gotchas from authoring Claude Code plugins. Includes a reference skill for the YAML parse trap on slash-command argument-hint values.

Hard-won gotchas from CI pipelines and build environments: GitHub Actions, Copilot CLI headless runs, cron block generation, Docker on Apple Silicon, and TypeScript monorepo scaffolding.

Extracts reusable skills from Claude Code sessions at session end. Analyzes for non-obvious discoveries and creates SKILL.md files.

Development workflow tools — README generation and review workflows.

Design-and-build process skills: layered design docs, C4 architecture artifacts, Diataxis documentation, README authoring, a milestone execution harness, positive-framing spec guidance, and a schema drift playbook.

Hard-won gotchas from real git and gh workflows: PR create/merge quirks, worktree races, rebase in non-interactive shells, squash-merge diagnosis, and deadlock recovery via plumbing.

Operational gotchas and references for third-party integrations driven from Claude Code — Cloudflare Workers, Grafana, Notion, OpenPencil, Playwright, and MCP file transport.

iTerm2 terminal job dispatcher and session controller — manages terminal sessions, dispatches jobs, and monitors their progress.

iTerm2 utilities for Claude Code — pane and session management.

Convert markdown directories to PDF with Mermaid diagram rendering.

Persistent memory system using Obsidian vault. Teaches Claude when/how to use Obsidian for storing decisions, learnings, workflow facts, and retrieving context.

Automated PR monitoring with Stop hook that auto-resumes Claude Code when new commits are detected.

Hard-won gotchas from the Python toolchain: importlib class identity, pip/PyPI name collisions, mock.patch.object vs __slots__, ruff on notebooks, and uv in fresh worktrees.

AIDE²-style recursive self-improvement loop: an outer agent iteratively rewrites an inner tree-search research agent, keeping a rewrite only when it beats the incumbent on private held-out scores under a fixed budget across an optimization / ML-engineering / harness-engineering battery. Standalone /rsi:autoresearch solves tasks by tree search; /rsi:init + /rsi:step + /rsi:run drive the outer loop.

Semantic search over Obsidian notes using LanceDB and sentence-transformers for local embeddings.

Mine Claude Code sessions for knowledge extraction. Process session history to identify learnings, patterns, and improvement opportunities.

Skills for verifying that tests actually test something: a dependency-free harness for unit-testing Workflow scripts, mutation-verification for renderer tests, real-browser checks for Streamlit UIs, and zsh traps that fake out verification blocks.