SKILL.md Walkthrough -- Build, Test &
Day 60 made the case that SKILL.md is the new app-store unit. Today is the hands-on follow-up: pick a narrow workflow, ship one folder, distribute to four marketplaces, monetise per call -- all in under 90 minutes. This is the playbook the Anthropic 17-skill repo, the Microsoft / OpenAI / Atlassian / Figma / Canva / Stripe / Notion / Zapier adoptions, and Google's Gemini Spark Skills (launched May 19 at I/O) all converge on.
Governance, IP, Litigation) + 20+ MCP connectors (Harvey, Solve Intelligence, Definely, DocuSign, Ironclad, Thomson
Reuters, Midpage, Trellis, Legal Data Hunter, Descrybe, Free Law Project) -- the largest single skill-plus-connector drop in 2026. Legal is now the most engaged Cowork knowledge-work function, validating the 'one skill, four channels' playbook at vertical scale. Free Law Project + Justice Technology Association partnerships extend reach to access-to-justice users. Lesson for builders: vertical plugin bundles ship faster than horizontal frameworks and capture 6-figure enterprise deals. practice-area plugins shipped MCP connectors live embedded AI agents inside the suite 1 · The 6-Step Build (90 minutes, end-to-end) The contract-redline skill is the canonical example because it touches three p
1 · The 6-Step Build (90 minutes, end-to-end)
The contract-redline skill is the canonical example because it touches three primitives most enterprise agents need: document parsing, instruction-following, and structured output. Pick a workflow you do 3+ times a week -- that's the right scope for a v1 skill. Write the SKILL.md first, instruments and monetisation come last. Total clock time on a fresh laptop: 75-90 minutes.
| Step | Action | Tool / Output |
|---|---|---|
| 1. Scope (5 min) | Pick one narrow workflow you repeat weekly. Write a 1-sentence trigger ('When user uploads a .docx contract and asks for redline...') | Notepad |
| 2. Author SKILL.md (20 min) | YAML frontmatter: name, description (must read like a trigger), allowed-tools (the trust boundary), model preferences. Markdown body: do/dont examples + numbered steps + script refs. | VS Code |
| 3. Test locally (15 min) | Drop folder into ~/.claude/skills/. Invoke via Claude Code. Build 20-pair golden set in Promptfoo. Block PR on >2% regression. | Claude Code CLI + Promptfoo |
2 · SKILL.md Anatomy -- The Frontmatter Is the Trust Boundary
A skill is just a folder with SKILL.md at its root. YAML frontmatter declares name, description, allowed-tools, and model preferences -- the agent loads the skill only when description matches user intent, preserving context budget (same lazy-load pattern that made MCP scale in 2025). Critically: the allowed-tools list IS the trust boundary. Any tool not on that list cannot be called when this skill is active. This is the single most important governance lever -- get it wrong and you've shipped a privilege-escalation vector to every marketplace simultaneously.
| Step | Action | Tool / Output |
|---|---|---|
| 4. Wrap as MCP server (15 min) | FastMCP 3.0 - one @tool decorator + reuse SKILL.md body as system prompt. ~60 lines of Python. Add Zuplo or eco.com x402 middleware for monetisation. | FastMCP + Zuplo |
| 5. Register on A2A (10 min) | Sign Agent Card with DID, publish to a2a-registry (150+ orgs already in production, governed by Linux Foundation Agentic AI Foundation). | Linux Foundation registry |
| 6. Publish to all 4 channels (10 min) | PR to anthropics/skills + zip to GPT Store + Fluora MCP marketplace listing + Gemini Spark Skills developer console. | One artefact, four channels |
3 · Testing & Eval Loop -- Promptfoo + OTEL gen_ai From Day
The Day 50 closed eval loop applies fully to skills. Build a 20-pair golden set from real prod usage (not synthetic) before publishing. Wire Promptfoo into CI as a regression gate -- any >2% drop blocks merge. Add OTEL gen_ai spans on every tool call so Phoenix / Langfuse / Galileo Signals can flag failure modes weekly and auto-promote them into the golden set. This is what turns a skill from a one-off demo into a self-healing product. Datasets become living artefacts.
| Step | Action | Tool / Output |
|---|---|---|
| 4. Wrap as MCP server (15 min) | FastMCP 3.0 - one @tool decorator + reuse SKILL.md body as system prompt. ~60 lines of Python. Add Zuplo or eco.com x402 middleware for monetisation. | FastMCP + Zuplo |
| 5. Register on A2A (10 min) | Sign Agent Card with DID, publish to a2a-registry (150+ orgs already in production, governed by Linux Foundation Agentic AI Foundation). | Linux Foundation registry |
| 6. Publish to all 4 channels (10 min) | PR to anthropics/skills + zip to GPT Store + Fluora MCP marketplace listing + Gemini Spark Skills developer console. | One artefact, four channels |
4 · Distribute & Monetise -- Four Channels, One Artefact
Each marketplace has different economics. Anthropic Skills (free open standard, drives Cowork stickiness, Claude.ai Directory unified at claude.ai/directory March 31). GPT Store (70/30 split, 4M+ Codex weekly users + 900M ChatGPT WAU). Fluora MCP marketplace + x402 USDC on Base (sub-cent per-call, no signup, $600M+ in 500K+ active wallets, Coinbase reports 165M transactions across 69K active agents as of April). Gemini Spark Skills (announced at Google I/O May 19, Gemini app reach 3B+ Android + Workspace users, Spark Skills inside the Gemini app with Tasks + Schedules orchestration, Android Halo
| Layer | Tool | Trigger |
|---|---|---|
| Auto-promotion | LangSmith Insights Agent or Galileo Signals | Cluster failures weekly, draft new golden-set entries |
5 · Governance -- Skills Are Privilege-Escalation Vectors
Every skill ships an allowed-tools list -- that list IS the trust boundary. Treat community skills like npm packages from unknown authors: sandbox them, scope OAuth per skill, and log every skill load to a WORM audit store. The npm / PyPI supply-chain attack pattern WILL hit Skills marketplaces in H2 2026 -- likely a typo-squatted contract-redline-pro skill that exfiltrates docs. Reputation gating (downloads + maintainer DID + Linux Foundation signed Agent Card) is the only durable defence. EU AI Act enforcement at T-71 days (Aug 2) means every skill load on a high-risk system needs Annex III audit
| Layer | Tool | Trigger |
|---|---|---|
| Auto-promotion | LangSmith Insights Agent or Galileo Signals | Cluster failures weekly, draft new golden-set entries |
6 · Breaking This Week
Anthropic Legal Plugin Suite (May 12) -- 12 practice-area plugins + 20+ connectors, largest single vertical skill drop in 2026. Legal already the most engaged Cowork knowledge-work function. Google Gemini Spark + 3.5 Flash (May 19 I/O) -- Spark Skills inside the Gemini app, Android Halo UI surface coming later this year, Spark runs on dedicated Google Cloud VMs (24/7 background). Google Antigravity 2.0 -- community benchmarks claim 12× speed vs Claude Code on some workflows but Annex III compliance docs missing. xAI Grok Build (May 20) -- coding agent at $30/mo. Walmart Sparky AI assistant 40% smarter, weekly users 2×. Netflix hands ad-buying to AI agents. Klarna shopping engine inside ChatGPT (100M products wired). ChatGPT now connects to financial accounts via Plaid for Pro users. Anthropic in talks for $30-50B raise at ~$950B valuation (ARR $44B = 80× Q1 2025). EU AI Act T-71 days, 78% enterprises NOT on track per Secure Privacy survey.
MARKET SIGNAL
Model performance is compressing (<3% across Opus 4.7 / GPT-5.5 / Gemini 3.1 Pro / Qwen3.6+). The next $50B+ in agentic value is captured by distribution and skills ecosystem and governance evidence -- not by another point on a benchmark. Anthropic plays the open-protocol / npm card. Google plays the OS-surface card via Gemini Spark + Android Halo. Microsoft plays the enterprise control-plane card via Agent 365. The asymmetry window closes in 2027 when revenue shares formalise. Build skills now while the marketplaces are still hungry for inventory.
VIRAL APP -- Anthropic Legal Plugin Suite (May 12, 2026)
12 practice-area plugins (Commercial, Corporate/M&A diligence, Employment, Privacy, Product, Regulatory, AI Governance, IP, Litigation) + 20+ MCP connectors (Harvey, Solve Intelligence, Definely, DocuSign, Ironclad, Thomson Reuters, Midpage, Trellis, Legal Data Hunter, Descrybe, Free Law Project) -- the largest single skill-plus-connector drop in
| Step | Action | Tool / Output |
|---|---|---|
| 1. Scope (5 min) | Pick one narrow workflow you repeat weekly. Write a 1-sentence trigger ('When user uploads a .docx contract and asks for redline...') | Notepad |
| 2. Author SKILL.md (20 min) | YAML frontmatter: name, description (must read like a trigger), allowed-tools (the trust boundary), model preferences. Markdown body: do/dont examples + numbered steps + script refs. | VS Code |
| 3. Test locally (15 min) | Drop folder into ~/.claude/skills/. Invoke via Claude Code. Build 20-pair golden set in Promptfoo. Block PR on >2% regression. | Claude Code CLI + Promptfoo |
Model performance is compressing (<3% across Opus 4.7 / GPT-5.5 / Gemini 3.1 Pro / Qwen3.6+). The next $50B+ in agentic value is captured by distribution and skills ecosystem and governance evidence -- not by another point on a benchmark. Anthropic plays the open-protocol / npm card.
<b>Ship one skill this weekend.</b> Pick a workflow you do 3+ times a week. SKILL.md it. Test on a 20-pair golden set in Promptfoo. Total time: 90 minutes. Repeat next weekend with workflow #2. <b>Frontmatter discipline = governance.</b> Treat the allowed-tools list as the trust boundary. Scope every OAuth token to exactly that list. Reject any tool call outside it at host level -- not at app level. <b>Wrap as MCP + register on A2A from day one.</b> Same SKILL.md body becomes the MCP server's system prompt (~60 lines FastMCP). DID-signed Agent Card on Linux Foundation a2a-registry is your supply-chain defence. <b>Publish to all four channels in parallel.</b> Anthropic Skills (free) + GPT Store (70/30) + Fluora (per-call x402) + Gemini Spark Skills (developer console). One artefact, four revenue surfaces. <b>Wire OTEL gen_ai spans before publishing.</b> Phoenix or Langfuse on day one. Galileo Signals or LangSmith Insights Agent finds new failure modes weekly and auto-promotes them into the golden set -- closed self-healing loop from <b>Tomorrow:</b> Day 62 (2026-05-24): The Agent Marketplace Wars -- economics, take rates, and discovery algorithms across Anthropic Skills / GPT Store / Fluora / Gemini Spark Skills / Microsoft Agent Store. Which channel actually pays out, which captures the long tail, and where to bet your second skill. varunsingla.com · Daily learning brief · Built with Claude in Cowork mode