Four Claude Code skills extracted from shipping production AI products end-to-end with the agentic coding loop. Each skill codifies a discipline that survives contact with paying users.
| Skill | What it does | Source pattern |
|---|---|---|
mcp-directory-readiness |
Audits any MCP server against the Anthropic MCP Directory submission requirements; generates gap list + pre-filled form data | Anthropic's published submission flow + my own mcp-butterfly submission run |
anthropic-api-resilience |
Drops round-robin keys + 1s→5min exponential backoff + circuit breaker into any Anthropic SDK client | Production reliability pattern surfaced in Anthropic's internal RL Studio codebase |
evidence-gated-ci |
Scaffolds smoketest + e2e + artifact-verified-mutation CI gate — no external mutation ships without a captured response | Discipline I run across Tandem, Butterfly Security, LinguaMind, and mcp-butterfly |
multi-model-routing |
Cost-aware router with per-task-type model selection + per-call telemetry | Multi-model routing pattern from Mercor's interview scoring codebase |
Each skill is one folder under skills/ containing a SKILL.md with YAML frontmatter and supporting templates (TypeScript + Python drop-in code, reference docs, checklists).
Symlink any individual skill into your Claude Code skills directory:
ln -s "$(pwd)/skills/mcp-directory-readiness" ~/.claude/skills/mcp-directory-readinessThe skill auto-registers on the next Claude Code session start.
for s in skills/*/; do
ln -s "$(pwd)/$s" "$HOME/.claude/skills/$(basename $s)"
doneOnce installed, invoke any skill from Claude Code by name:
/skill mcp-directory-readiness
/skill anthropic-api-resilience
/skill evidence-gated-ci
/skill multi-model-routing
Or just describe the work — each skill's frontmatter lists the natural-language triggers ("audit my MCP server", "Claude API rate limit", "add CI to this project", "route between Claude and GPT").
I ship production AI products solo with Claude Code as the primary editor. These four skills are the disciplines that, when missing, broke something in front of real users.
mcp-directory-readiness— built while preparingmcp-butterflyfor the Anthropic MCP Directory. The mechanical checks (tool annotations, OAuth posture, HTTPS) account for ~80% of submission rejections and are entirely automatable.anthropic-api-resilience— built after watchingLinguaMindsilently drop user requests during a Claude API rate-limit window. Single-key clients fail badly at scale; round-robin + backoff converts it to invisible failover.evidence-gated-ci— codifies what I do in every project's CLAUDE.md so it's portable. The load-bearing layer is artifact-verified mutations: no claim of external state change ships without a captured HTTP response on disk. Prevents the class of agent-fabricated state changes that pass review because nobody asked the external system whether the mutation actually happened.multi-model-routing— built for the LinguaMind tutor (vocab extraction → Haiku, RAG synthesis → larger model). Routing by task-type with per-call cost telemetry compounds savings as DAU scales.
- Tandem — universal ebook + audiobook reader, live on iOS / iPadOS / macOS, visionOS in App Review. SwiftUI across four Apple platforms.
- Butterfly Security — paying-customer SaaS for identity-infrastructure disaster recovery. Listed in the Okta Integration Network as an API Service Integration.
- LinguaMind — AI language coach on Cloudflare Pages + D1 + Workers AI + Claude Haiku.
- mcp-butterfly — remote streamable-HTTP MCP server on Cloudflare Workers + Durable Objects exposing Butterfly Security as Claude-callable tools.
Issues and PRs welcome. Each skill has its own templates and reference files in the skills/<skill-name>/ folder — keep the structure consistent if you're adding a new skill.
MIT — see LICENSE.