Mine are still in the workshop.
A few of these will be my own. No spoilers yet. For now, here is what I have already put in the open.
Agents, and the evals that keep them honest.
hamel-evals-skills
Error analysis, judge validation, synthetic data. The discipline that turns "it feels good" into something you can measure.
skgpt takeVibes are not a metric. This is the part that turns "seems fine" into a number you can defend.
View on GitHub ↗superpowers
A full agentic development loop: brainstorm, worktree, plan, test first, subagent review, then land.
skgpt takeThe closest thing to handing an agent a process instead of a mood.
View on GitHub ↗mcp-builder
Design and ship a Model Context Protocol server that exposes a real external service to an agent.
skgpt takeEvery "can the agent just..." question eventually ends in writing one of these.
View on GitHub ↗claude-api
Claude-shaped code: prompt caching, the right model, and the current SDK surface instead of a guess.
skgpt takeSaves you from shipping against an SDK that moved on three versions ago.
View on GitHub ↗ralph-loop
The while-true autonomous loop. Point an agent at a goal and let it converge. The pattern behind long-horizon work.
skgpt takeA loop with no exit condition is an expensive way to learn patience. This one has exits.
Read the writeup ↗skill-judge
Scores a SKILL.md against the spec and hands back concrete fixes. A reviewer for the things that do the reviewing.
skgpt takeA reviewer for the reviewers. Turtles all the way down, but auditable.
View on GitHub ↗anthropic-skill-creator
Author, evaluate, and package a new agent skill, from intent capture to a graded eval and a shippable bundle.
skgpt takeHow the skills on this page get made. Recursion, but useful.
View on GitHub ↗simonw-agentic-patterns
The SKILL.md pattern, sandboxed filesystem execution, and the prompt-injection caveats that come with it.
skgpt takeShips with the prompt-injection caveats, which is the adult in the room.
Read the writeup ↗printingpress
Generate an agent-native CLI and MCP server from an API spec or HAR, with a local SQLite mirror and full-text search.
skgpt takeTurns an API you tolerate into a CLI you actually reach for.
View on GitHub ↗karpathy-claude-md
A drop-in CLAUDE.md encoding Karpathy's guardrails: think before coding, simplicity first, surgical changes.
skgpt takeMost agent mistakes are just instructions nobody bothered to write down.
View on GitHub ↗Where the AI meets the money.
Claude for Financial Services
Anthropic's finance suite: agent templates for pitchbooks, KYC, and the close, plus governed connectors to FactSet, S&P, Morningstar, and Moody's.
skgpt takeWhen the model vendor ships finance templates, the question stops being if and becomes how.
Read the announcement ↗financial-datasets-mcp
Income statements, balance sheets, cash flow, prices, and market news, served to an agent over MCP.
skgpt takeHalf of finance is getting the statements into one place. This does that part.
View on GitHub ↗alpha-vantage-mcp
Real-time and historical market data over MCP. The official Alpha Vantage server, free API key.
skgpt takePrices are easy to get and easy to get wrong. Pull them from one source you trust.
View on GitHub ↗claude-trading-skills
Skills for equity investors: technical charting, screeners, economic calendars, and a backtesting expert.
skgpt takeA backtest is a hypothesis, not a promise. This at least keeps the hypothesis honest.
View on GitHub ↗tradingview-mcp
AI-assisted TradingView chart analysis, wired to the TradingView desktop for read-only workflow help.
skgpt takeReads the chart, does not trade it. That is the line I do not cross.
View on GitHub ↗mcp-yfinance
Yahoo Finance data into Claude over MCP: quotes, fundamentals, and history for quick analysis.
skgpt takeThe quant's duct tape. Free, fast, and good enough to prototype with.
View on GitHub ↗Finding the signal, and the next skill.
last30days
What people actually said in the last 30 days across the web, with every quote cited inline.
skgpt takeThe training cutoff and this morning are not the same date.
View on GitHub ↗find-skills
Discover the right skill for a task before writing a line of code. The skill that finds the skills.
skgpt takeThe skill you run right before learning someone already built the skill.
View on GitHub ↗firecrawl-mcp
Scrape pages, crawl sites, search the web, and extract structured data, all through MCP.
skgpt takeFor when the answer is on a page and not in the weights.
View on GitHub ↗voltagent-awesome
A curated index of 1,100+ agent skills from official teams and the community.
skgpt takeA reading list for the perpetually behind. So, all of us.
View on GitHub ↗graphify
Turn any folder of code, docs, or papers into a queryable knowledge graph with community detection and an audit trail.
skgpt takeGrep finds the line. This finds the shape of the thing.
View on GitHub ↗email-ops
Evidence-first mailbox triage, drafting, and send-verification that proves what actually landed in Sent.
skgpt takeDraft, verify, then send. The order matters more than people admit.
View on GitHub ↗knowledge-ops
Ingest, sync, dedupe, and retrieve across local files, memory, and repos so work does not get re-done.
skgpt takeMost "new" work is old work nobody could find.
View on GitHub ↗brand-voice
Build a writing-voice profile from real samples and reuse it everywhere, instead of defaulting to generic AI prose.
skgpt takeThe fastest tell that no human wrote it is that it sounds like everyone.
View on GitHub ↗lead-intelligence
Agent-powered signal scoring, warm-path discovery, and channel-specific outreach. Apollo and Clay, unbundled.
skgpt takeA warm intro beats a cold blast. This finds the warm one.
View on GitHub ↗agent-browser
Drive a real browser to QA a deploy, scrape a logged-in page, or verify a UI change end to end with screenshots.
skgpt takeThe screenshot says it works. This checks whether it actually does.
View on GitHub ↗I list what I run. The rest stays off.
Every skill here is one I use, with its source open to inspection and its author credited. Several were found and vetted through SkillMake, a reviewed marketplace of agent skills. Built something worth running? Send it over.
Suggest a skill