Infrastructure
Claude Code Skills
Over forty handcrafted skill templates that turn a general-purpose AI assistant into a team of specialists, each with its own methodology, quality standards, and domain knowledge baked into a single markdown file.
In Plain English
These are instruction files that teach Claude Code (an AI programming assistant that runs in your terminal) how to do specific jobs. Instead of explaining the same complex task every time, you type a short command like /webforge or /reflect, and the AI loads everything it needs to become an expert at that particular task. Think of them as recipes that turn a general-purpose AI into a specialist on demand.
Problem
Large language models are generalists by design. Ask Claude to build a website and it will produce something functional but generic. Ask it to write shadow work journal prompts rooted in Jungian psychology and Internal Family Systems therapy, and it will produce something that reads like a Wikipedia summary. Ask it to review code with the rigor of a senior engineer, and it will find surface-level issues but miss the architectural concerns. The knowledge is in there somewhere, but without structure and methodology, it comes out diluted.
Claude Code's skill system solves this with a deceptively simple mechanism: a markdown file that loads into context when a slash command is triggered. But the simplicity of the delivery mechanism hides the complexity of what goes inside. A well-designed skill is not just a prompt. It is a complete methodology document that includes decision trees, quality rubrics, anti-pattern lists, tool permissions, and step-by-step workflows. It transforms the AI from "capable of doing the task" to "has a proven process for doing the task well."
This collection of forty-plus skills covers an unusually wide range: from engineering workflows like /compound (plan, work, review, learn) and /webforge (multi-agent website building with six specialist roles) to deeply personal ones like /psychology (Jungian shadow work structure with somatic awareness prompts) and /humanizer (detecting and removing thirty-plus AI writing patterns). Each skill was authored, tested in real use, and refined through the /reflect learning loop until it consistently produced high-quality output for its domain.
Architecture
Key Features
Multi-Agent Orchestration Skills
/webforge, /compound, /director
The most sophisticated skills do not just instruct the AI to do one thing. They define entire multi-step workflows with multiple specialist personas. The /webforge skill, for example, instantiates six virtual roles: a design architect who establishes visual identity, a performance engineer who optimizes Core Web Vitals, an accessibility auditor, an SEO specialist, a content strategist, and an anti-AI polish inspector who catches the telltale signs of generated code. The /compound skill orchestrates the Plan-Work-Review-Learn loop that structures all significant development tasks.
Deep Domain Knowledge
/psychology, /integration-guide, /obsidian-writer
Several skills encode genuinely specialized knowledge that the base model handles loosely at best. The /psychology skill understands Internal Family Systems therapy parts, Jungian archetypes, somatic body awareness, and the structure of shadow work entries including trigger types, intensity scales, and body location tracking. The /obsidian-writer knows the vault's exact folder hierarchy, frontmatter schemas for each content type, the tag taxonomy, and Obsidian-specific formatting like callouts and DataviewJS queries. These are not summaries. They are working reference documents.
Self-Improving Methodology
/reflect, /iterate, /challenge
A core set of skills exists to make the other skills better over time. The /reflect skill runs after any significant task, extracting structured learnings: what approach was used, what worked, what failed, what was surprising. These get recorded as JSONL entries in a learnings store and fed to the memory system. The /iterate skill applies incremental refinement to existing work. The /challenge skill deliberately stress-tests assumptions, pushing the AI to find holes in its own plans before they become bugs in production.
Quality Control and Anti-Patterns
/humanizer, /review, /harden
Every skill in the collection includes an anti-pattern section, but some are built entirely around quality enforcement. The /humanizer skill carries a tiered kill-list of over thirty AI writing patterns (words like "delve," "tapestry," and "paramount" that appear ten to one hundred times more often in AI text than human text) with natural replacements for each. The /review skill performs multi-pass code analysis covering architecture, security, performance, and maintainability. The /harden skill locks down applications against common attack vectors with a structured security audit methodology.
How It Works
Skill Authoring
Each skill starts as a SKILL.md file with a YAML frontmatter header. The header declares the skill name, a trigger description, and an explicit allowed-tools list that constrains which Claude Code tools the skill can call. A /psychology skill has no business calling Bash, so it is restricted to Read, Write, and Obsidian MCP tools. A /browser-agent skill needs Bash to launch the automation script but nothing else. This principle-of-least-privilege approach keeps skills focused and prevents side effects.
Trigger and Loading
When a user types a slash command like /webforge, Claude Code scans the skills directories for a matching name and loads the entire SKILL.md into context. The skill's content, methodology, decision trees, and quality rubrics are all instantly available to the model. Some skills also auto-trigger based on keyword detection: saying "shadow work" loads the psychology writer, saying "optimize" loads the system tuner, saying "budget" loads the finance advisor. The /director skill handles this routing for ambiguous cases.
Methodology Execution
Once loaded, the skill does not just provide knowledge. It provides a step-by-step workflow that Claude follows sequentially. The /compound skill, for instance, forces the AI through four phases: planning (80% of effort, including memory recall and learnings lookup), working (executing the plan in small verifiable steps), reviewing (a fresh-context subagent audits the work), and compounding (extracting and recording learnings). Each phase has explicit entry and exit criteria. The AI cannot skip ahead because the workflow is written as imperative instructions, not suggestions.
Packaging and Distribution
Skills that live in the .claude/skills/ directory are automatically available to Claude Code on every session. For Claude.ai web sessions, a separate packaging pipeline bundles skills into .skill files that can be drag-and-dropped into a conversation. The /skill-creator skill automates this entire lifecycle: scaffolding a new skill from a template, validating its structure against the required schema, and packaging it for distribution. A Python validation script checks for required sections, proper YAML frontmatter, and tool permission consistency.
Continuous Refinement
Skills are living documents that evolve through use. Every time the /reflect skill runs after a task that used another skill, it checks whether the methodology held up or whether adjustments are needed. Anti-patterns discovered in practice get added to the skill's kill-list. New decision tree branches get added when edge cases emerge. The learnings store feeds back into skills through the /compound loop, creating a system where every completed task makes the next invocation of that skill slightly more capable. Skills are versioned alongside the rest of the codebase.
Tech Stack
Format
Markdown with YAML frontmatter
Runtime
Claude Code CLI skill loader
Packaging
.skill bundles for Claude.ai
Tooling
Python (init, validate, package scripts)
Storage
JSONL learnings store
Integration
Memory system, MCP tools, subagents