Claude Code : Agent Skills, Agent Loop, Code Agents, MCP

Explore cloud code architecture from the user interface to external services. Learn how a Google login feature is planned, authenticated via OAuth, and integrated across backend and frontend.

Learn how plot skills prevent repetitive instructions, eliminate manual context pasting, and ensure consistent institutional knowledge with specialized domain expertise, a lightweight token footprint, and on-demand context loading across sessions.

Identify whether tasks require real-time external data, autonomous multi-step execution, or repeatable domain knowledge to decide if skills, MCP, or agents are best.

Install skills on Cloud AI by enabling code execution, adding or selecting skills (including MCP Builder), and creating or uploading skills, then toggle them on to run locally.

Install skills on CloudCore and Cloud API by creating a .cloud directory with a skills file at project or user level, enabling team sharing and automatic discovery via API v1.

Create a valid skills markdown file from scratch and master Yammer front matter rules, install skills in Cloud Code or Cloud AI, and verify them with a test prompt.

Hands-on with Claude Code skills: reload a meeting notes skill, format text as meeting notes, and generate a meeting summary, key decisions, action items, and next steps in cloud code.

Understand the skills architecture and on-demand loading in cloud, where the VM file system's skills directory feeds the context window based on user requests, loading only the requested skill.

Cloud loads only what's needed by matching pdf text to the pdf processing skill's description, loading the relevant instruction and references, illustrating progressive disclosure for efficient skill execution.

Explore progressive disclosure as a design principle for dynamic skills loading: guide with references, domain organization, and conditional scripts to enable plan, validate, execute workflows.

master progressive disclosure by choosing right file structure via two patterns: guide with references and domain-specific organization; follow five rules, including under 500 lines, one level deep, and descriptive names.

Plan the skill before coding to define its domain, scope boundary, and necessary decisions for code review, including degree of freedom and pattern choices (template, example, workflow).

Explore code review fundamentals with a structured walkthrough that defines skill naming, description, calibration levels, review sections, and practical examples for quality auditing.

Execute a code review on a Python dbconnect file that reads a key-value config, connects to a database, and outputs a structured review with severity, strengths, and recommended next steps.

Master the skills anatomy and the YAML front matter, with name and description fields, and learn how the two zones—front matter and body—drive Claude's discovery with triggers.

Apply the design principle of the right degree of freedom for cloud skills, using a terrain analogy to decide high, medium, and low freedom based on the cost of deviation.

Explore the three design patterns—template, example, and workflow patterns—and how to structure skills files with fixed sections, input-output pairs, and a copyable checklist.

Describe three loading levels for skills, from startup front matter to domain expertise. Load level 1 at startup, level 2 on demand, level 3 as needed, with token optimization.

Explains how to invoke Anthropic skills via API, outlining required beta headers, code execution, skills references, and files API for downloading PowerPoint, Excel, Doc, or PDF outputs.

Explore how slash commands in Cloud Code automate repetitive tasks with reusable markdown templates, enabling project-based and user-scoped workflows, built-in and custom commands, for consistent team collaboration.

Master the slash command structure by learning the forward slash, command name, and positional arguments, illustrated with fix issue 123 high priority as the example.

Explore using slash commands with multiple inputs by leveraging positional parameters to pass arguments such as PR 121, medium priority, and John as the assignee, illustrated via ReviewPR.md.