I Know Kung Fu

One of the most iconic scenes in The Matrix is when Neo downloads martial arts directly into his brain. Seconds later: "I know Kung Fu."

We're entering that era. But instead of martial arts, we're loading capabilities. And one of the most underrated tools in this new world is something called Agent Skills — and SKILL.md.

What are Agent Skills?

Agent Skills are a simple, open format for giving AI agents new capabilities through structured, reusable expertise. At their core, a skill is just a folder containing a SKILL.md file with clear metadata and step-by-step instructions for performing a specific task.

Skills can also include scripts, templates, references, and other assets, allowing teams to bundle procedural knowledge, domain expertise, and repeatable workflows into portable packages.

Instead of relying purely on generic model intelligence, skills give agents access to company-, team-, or user-specific context that they can load when needed. They work through progressive disclosure: agents first read only a skill's name and description to know when it's relevant, then load the full instructions when a matching task appears, and finally execute the workflow.

This keeps agents efficient while enabling deeper specialization on demand.

Why Context Matters

Large Language Models have limited context windows (although these are growing massively with each new iterative launch), and for them to be useful, user interfaces like chat interfaces and agentic tools ingest system prompts and other useful content.

But context bloat is a real problem. The more instructions, examples, guardrails, and background you stuff into the window, the more you dilute focus and increase cost.

So the idea is to be useful and lightweight. Instead of loading everything all the time, you structure capabilities so they can be discovered and activated only when needed. That keeps agents efficient, reduces noise, and preserves high-quality reasoning within a constrained context window.

Domain Expertise in the World of Generative AI

In the context of domain expertise, there are three different ways to make AI "smarter" about a specific field:

Fine-tuning means retraining a model on specialized data so it internalizes a domain's patterns (e.g., legal contracts, medical notes, financial reports). The knowledge becomes baked into the model's weights. It's powerful but heavier, slower to update, and less flexible.

RAG (Retrieval-Augmented Generation) means the model doesn't memorize the domain, it looks things up at runtime. It retrieves relevant documents from a knowledge base and uses them to answer accurately. This is great for up-to-date, factual, organization-specific knowledge.

Skills sit slightly differently. They don't just give the model knowledge, they give it structured procedural capability. A skill defines how to do something in a domain (workflows, rules, constraints, steps), and the agent loads it only when needed.

So instead of just knowing finance, it knows how to execute a financial analysis process your way. In other words, it turns vague claims like "I'm good at strategy" into operational intelligence. It allows AI systems to understand and execute based on defined capability boundaries.

The Difference

It's the difference between:

"Help me with marketing."

and

"Apply B2B enterprise positioning frameworks with pricing psychology and GTM sequencing for emerging markets."

One is generic. The other is downloadable Kung Fu.

---

Why This Matters

We are moving from simply prompting models to defining structured capability layers. AI will not just respond to prompts; it will operate through clearly defined skills.

Companies that formalize their skills, both internally and digitally, will compound their leverage. Individuals who define their skills precisely will become more capable.

This isn't about replacing people — it's about compressing expertise into structured intelligence that can be deployed, scaled, and activated on demand.

---

An Interesting Observation

I wanted to create an equity research skill for one of my agents. Using one LLM to help "train" another LLM seems like the only rational choice.

To use this as an opportunity to grade some of these SOTA models on a practical task, I gave the same prompt to three models, Claude Sonnet 4.6, ChatGPT 5.2 and Gemini 3:

"I need to create a SKILL.md file for an equity research agent. I need the agent to be able to do equity research on a stock... look at call transcripts, earnings reports, EDGAR filings, and news articles and anything else you can think of."

The difference in outputs was striking.

Claude blew it out of the park.

It delivered a structured, procedural, step-by-step framework: identify the company, pull EDGAR filings, analyze earnings reports, review call transcripts, scan news coverage, synthesize outputs — and for each step it referenced relevant digital resources.

It was properly formatted in clean markdown, aligned to the SKILL.md structure, and felt operational. 9/10.

ChatGPT was a mediocre generalist.

The response was coherent and useful at a high level, but it lacked concrete digital resources and procedural depth. It explained what equity research involves, but didn't fully translate that into a structured, executable skill. 6/10.

Gemini was the weakest.

The output fell below expectations. Gemini seemed to be more focused on generating images than producing structured data. 4/10.

---

The Takeaway

The takeaway isn't that one model "wins" universally. I've had vastly different experiences with each of them. In this instance, Claude is the clear winner.

For reference, here is the SKILL.md I went with:

---

name: equity-research description: > Comprehensive equity research agent that analyzes stocks using earnings call transcripts, earnings reports, SEC/EDGAR filings, and news articles to produce investment-grade research.

Use this skill whenever a user asks to research a stock, analyze a company, evaluate an investment, review earnings, summarize 10-K or 10-Q filings, pull call transcripts, assess valuation, or generate an equity research report.

Trigger even for partial requests like "what do you think about NVDA?" or "help me understand Apple's latest earnings" — any question about a public company's fundamentals, outlook, or recent developments should use this skill.

Equity Research Agent

A skill for producing thorough, investment-grade equity research on publicly traded companies using primary sources: SEC filings, earnings call transcripts, earnings reports, and news.

Workflow Overview

When asked to research a stock, follow this pipeline in order. Adapt depth to the user's request (quick take vs. full report).

1. Identify the company & ticker
2. Pull recent EDGAR filings (10-K, 10-Q, 8-K)
3. Find earnings call transcripts (last 1–4 quarters)
4. Review latest earnings releases / press releases
5. Search recent news & analyst coverage
6. Synthesize into structured research output

Step 1 — Identify the Company

Confirm the correct ticker and exchange before proceeding. If the user is ambiguous (e.g., "research Apple"), confirm AAPL on NASDAQ. Note the fiscal year end date, as it affects which filings are most current.

Step 2 — EDGAR Filings

Base URL: https://www.sec.gov/cgi-bin/browse-edgar

Key URLs to fetch:

EDGAR full-text search: https://efts.sec.gov/LATEST/search-index?q="%22{COMPANY_NAME}%22"&forms=10-K

Always extract from filings:

• Revenue, gross margin, operating margin, net income (last 3 years minimum)
• Free cash flow and capex
• Debt levels, cash, and liquidity position
• Management guidance (if provided)
• Key risk factors specific to the company (not just boilerplate)
• Segment breakdown if the company has multiple business lines

Step 3 — Earnings Call Transcripts

Transcripts reveal management tone, forward-looking commentary, and analyst concerns that don't appear in the formal filings.

Sources to search (in priority order):

1. Seeking Alpha — https://seekingalpha.com/symbol/{TICKER}/earnings/transcripts
2. Motley Fool — https://www.fool.com/earnings-call-transcripts/
3. The Motley Fool Transcripts search — web search "{COMPANY} earnings call transcript Q{N} {YEAR}"
4. Investor Relations page — many companies post transcripts directly; search "{COMPANY} investor relations earnings transcript"

Pull the most recent 2 quarters minimum. For a full report, pull 4 quarters.

Key things to extract from transcripts:

• Management prepared remarks: Tone, strategic priorities, what they emphasize
• Guidance: Revenue, EPS, margin guidance for next quarter/year
• Analyst Q&A: What questions are analysts asking? What did management dodge or hedge?
• Key metrics called out: GMV, ARR, NRR, DAU, backlog — whatever the company tracks
• Changes quarter-over-quarter: Did management change their language on key topics?
• Red flags: Defensive answers, excessive caveats, sudden change in CFO/CEO tone

Step 4 — Earnings Reports & Press Releases

Beyond the formal 10-Q/10-K, earnings press releases (typically filed as 8-K exhibits) contain:

• Headline revenue and EPS vs. consensus
• Non-GAAP metrics (adjusted EBITDA, adjusted EPS, free cash flow)
• Management commentary
• Updated guidance tables

Search: "{COMPANY} Q{N} {YEAR} earnings results press release" or fetch directly from the IR page.

Always note: Beat/miss on revenue, EPS, and guidance raise/lower vs. prior quarter.

Step 5 — News & Analyst Coverage

Web searches to run: "{TICKER} analyst rating price target {YEAR}" "{COMPANY} news last 30 days" "{COMPANY} competitive landscape {YEAR}" "{COMPANY} {key product or segment} growth" "{COMPANY} insider buying selling {YEAR}"

Sources to prioritize:

• CNBC, Reuters, Bloomberg, WSJ for news
• Barron's, Seeking Alpha for analysis
• Yahoo Finance for quick consensus data: https://finance.yahoo.com/quote/{TICKER}/analysis

Extract:

• Current analyst consensus (Buy/Hold/Sell ratio, average price target)
• Recent rating changes
• Major recent news events (acquisitions, leadership changes, product launches, regulatory)
• Competitive dynamics and market share trends
• Macro factors affecting the sector

Step 6 — Synthesis & Output

Standard Output Structure

For a full equity research report, produce the following sections. Adjust for shorter requests (e.g., if user just wants "quick take", condense to 2–3 paragraphs).

🏢 Company Overview

• Business model in plain English (2–3 sentences)
• Primary revenue streams and their relative size
• Key competitive advantages / moat
• Geographic exposure

📊 Financial Snapshot

Present a clean table of the most recent data available:

📞 Earnings Call Highlights

• Key themes from management in the most recent call
• What analysts are focused on
• Forward guidance (exact numbers if given)
• Notable changes from prior quarter's tone or guidance

📁 Key Filing Findings

• Standout items from 10-K/10-Q (risks, MD&A, segment data)
• Any material 8-K events (acquisitions, guidance updates, leadership)
• Balance sheet health summary

📰 Recent News & Catalysts

• Top 3–5 news items from the last 30–90 days
• Upcoming catalysts (earnings date, product launches, investor day)

🎯 Analyst Sentiment

• Consensus rating and average price target vs. current price
• Bull case and bear case as articulated by analysts/management
• Recent rating changes

⚖️ Investment Considerations

Strengths:

• [List key strengths with supporting data]

Risks:

• [List key risks from filings and news — be specific, not generic]

Key Questions to Monitor:

• [What are the 2–3 things to watch in the next quarter?]

📌 Summary

2–3 paragraph synthesis of the overall picture. Present the facts and the range of views objectively. Do not make a buy/sell recommendation — instead, lay out the key factors an investor would need to weigh.

Research Principles

• Be source-specific: Always cite which filing, which quarter's transcript, or which news article you're drawing from. Don't blend sources without attribution.
• Use actual numbers: Avoid vague language like "revenue grew." Say "revenue grew 18% YoY to $24.7B in Q3 FY2024."
• Flag data gaps: If you couldn't find a transcript or a filing is unavailable, say so clearly rather than inferring.
• Be current: Prioritize the most recent 12 months of data. Note the date of the most recent filing or transcript used.
• Distinguish facts from interpretation: Separate what the documents say from what analysts/commentators interpret. Label opinions as such.
• Non-GAAP vs. GAAP: Always note when a metric is non-GAAP/adjusted and whether it differs materially from the GAAP equivalent.

Quick Reference — Common Metrics by Sector

Useful Base URLs

• SEC EDGAR Search: https://efts.sec.gov/LATEST/search-index?q="{TICKER}"&forms={FORM}
• SEC EDGAR Filings: https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK={TICKER}
• Yahoo Finance: https://finance.yahoo.com/quote/{TICKER}
• Yahoo Analysis: https://finance.yahoo.com/quote/{TICKER}/analysis
• Yahoo Financials: https://finance.yahoo.com/quote/{TICKER}/financials
• Macrotrends: https://www.macrotrends.net/stocks/charts/{TICKER}/{COMPANY}/revenue
• Seeking Alpha: https://seekingalpha.com/symbol/{TICKER}
• CNBC: https://quote.cnbc.com/quote-html-webservice/restQuote/symbolType/symbol?symbols={TICKER}&requestMethod=itv&noform=1&partnerId=2&fund=1&exthrs=1&output=json&events=1

Scope Calibration

Adjust research depth based on user request:

• Quick take (5 min): Latest earnings release + analyst ratings + 1 news item
• Standard (15 min): Above + 1 quarter of transcripts + key 10-Q findings
• Full report (30+ min): Above + 4 quarters of transcripts + 10-K deep dive + multiple news sources

Always confirm scope with the user if unclear.