Agent Architecture
Claude Skills: Teach AI Your Workflow
March 31, 2026 · 8 min read · Liam OBrien
Anthropic published a 33-page guide on Skills. This is the practical breakdown — what they are, how to build one, and what it looks like in practice.
What Is a Skill?
A Skill is a folder that teaches Claude how to do a specific task your way. You write it once. It's reusable across every session. Claude follows your instructions exactly — no drift, no improvisation.
Think of it like handing a new hire your playbook on day one. Except they never forget it.
Without a Skill
"Analyze this earnings call and look for contradictions with the 10-K filing. Focus on management tone vs legal disclosures. Check for new risk factors. Compare against last quarter. Format it like our internal research notes with the conviction score at the top..."
Re-explain every session. Inconsistent output. Easy to forget steps.
With a Skill
"Run the earnings analysis skill on NVDA."
Same methodology every time. Your research process encoded. Consistent output format.
What a Skill Looks Like
A skill is just a folder with a SKILL.md file and optional supporting files:
earnings-analysis/
├── SKILL.md # The instructions (required)
├── references/
│ ├── research-format.md # Your firm's note template
│ └── conviction-scale.md # How you score conviction
├── scripts/
│ └── fetch_filing.py # Pull 10-K data
└── assets/
└── output-template.md # Standard output structureExample: A Research Workflow Skill
Here's what the SKILL.md looks like for an earnings analysis skill. This is the file Claude reads to know what to do:
---
name: earnings-analysis
description: Analyze earnings calls against SEC filings
to find narrative divergence. Use when user says
"analyze earnings", "run earnings skill", or
provides a ticker with "10-K" or "transcript".
---
# Earnings Analysis
## Step 1: Gather Sources
- Pull latest earnings call transcript
- Pull most recent 10-K filing (Risk Factors, MD&A)
- Pull prior quarter's analysis if available
## Step 2: Narrative vs. Legal Comparison
- Extract management's key claims from transcript
- Cross-reference against 10-K legal disclosures
- Flag any claim where the filing contradicts
or hedges the narrative
## Step 3: Risk Factor Changes
- Compare current 10-K risk factors to prior filing
- Highlight NEW risks added this quarter
- Flag any removed or softened language
## Step 4: Output
- Format per references/research-format.md
- Conviction score per references/conviction-scale.md
- Lead with contradictions, then patterns, then confirmsThat's it. Claude reads this, follows the steps, uses the reference files for formatting, and produces consistent output every time. No re-explaining. No drift between sessions.
Three Types of Skills
Anthropic identifies three categories. Most firms will use all three:
Document & Asset Creation
Consistent, high-quality output every time. Research notes that follow your format. Client reports with your branding. Compliance memos with the right structure. No external tools needed — just Claude and your template.
Workflow Automation
Multi-step processes that follow your methodology. Research pipeline from idea to conviction score. Compliance review with validation gates. Onboarding checklist that never misses a step.
MCP Enhancement
Skills that orchestrate your connected tools. Pull data from AlphaSense, cross-reference SEC filings, draft the note, file it in your RMS — one workflow, multiple systems, coordinated by a single skill.
Five Architecture Patterns That Work
The guide documents five patterns for building production skills. Here's what each one does and when to use it:
Sequential Workflow
Use when: Multi-step processes in a specific order
Example: Research pipeline: gather sources → analyze → cross-reference → score conviction → format output
Multi-MCP Coordination
Use when: Workflows spanning multiple systems
Example: Pull from AlphaSense + check Bloomberg + draft in your template + file to SharePoint — one command
Iterative Refinement
Use when: Output quality improves with validation loops
Example: Draft analysis → fact-check gate → refine flagged claims → re-validate → publish. This is how ARIA works.
Context-Aware Tool Selection
Use when: Same goal, different tools depending on the situation
Example: Earnings data from AlphaSense for covered names, SEC EDGAR for new positions, manual transcript for private companies
Domain-Specific Intelligence
Use when: The skill carries institutional knowledge
Example: Your firm's investment philosophy, your compliance checklist, your research methodology — baked into the agent
Four Rules From Building in Production
Small skills that chain together beat one massive instruction set.
Each skill does one thing well. Chain them for complex workflows. Easier to test, version, and fix.
The description field is the most important line you write.
It's how Claude decides whether to load your skill. Include what it does AND when to trigger it. Vague descriptions = skill never loads.
Connect skills to your data through MCP servers.
A skill alone is a static playbook. Wire it to your CRM, research platform, or data feed and it becomes adaptive — executing differently based on real-time state.
Always hand-edit AI-generated skills.
First drafts are too verbose and too generic. The skill needs to reflect how your firm actually operates — not a generic approximation.
What's in the Full 33-Page Guide
This article covers the highlights. The full guide goes deeper on each:
Chapter 1
Fundamentals
Chapter 2
Planning & Design
Chapter 3
Testing & Iteration
Chapter 4
Distribution & Sharing
Chapter 5
Patterns & Troubleshooting
Chapter 6
Resources & References
Resources
Want to See Skills in Action?
We build custom agent architectures for investment managers — research agents, compliance automation, operational intelligence. All powered by composable skills connected to your firm's data.
Book a 30-Minute Scoping Call