Инженеры Anthropic раскрыли, как РЕАЛЬНО работают с Claude Code | 4 правила

Overview

The subtitles explain how Anthropic engineers (at an “AI COD/Claude Code Summit”) structure work with Claude Code, and why many users misunderstand it. The video emphasizes four rules for building and using skills—instead of repeatedly writing prompts.

Key Ideas / Technological Concepts

• Most work is repetitive, so re-writing prompts and context in every chat wastes both time and tokens.
• “Skills” are designed to be reusable units that persist across sessions, unlike one-off prompts.
• Claude Code is described as layered:
  1. Layer 1: the underlying model (e.g., Anthropic, OpenAI, Gemini).
  2. Layer 2: prompts + agents (manual “dialing” / selecting instructions).
  3. Layer 3: skills (reusable “applications” that survive beyond a chat).

The “4 Rules” (Product / Engineering Guidance)

Rule 1: Miss the “skills,” not the “prompts”

• Instead of creating a new prompt per task, package recurring tasks into skills.
• A skill behaves like a persistent folder/application, helping you avoid re-inventing style, tone, and context each session.

Rule 2: A skill has three layers, not just a prompt

A skill includes:

1. Description (header)

   • Short, specific text explaining when to use the skill.
   • Key claim: skill descriptions are loaded into context, so they must be concise and clear.

2. Instructions (main body)

   • Step-by-step rules, goals, constraints, and the expected output format.

3. Tools

   • Pre-built executable components (scripts, API wrappers, templates, terminal commands, etc.).
   • The agent runs these tools instead of repeatedly “re-thinking” and regenerating code via tokens.

Why tools matter: running code through tools is described as cheaper, faster, and more repeatable than regenerating code via token-based responses. The video claims many users neglect tools, resulting in weaker performance.

Rule 3: Compositional, not monolithic

• Avoid one huge “do-everything” skill/agent.
• Prefer 3–5 narrowly focused skills plus an orchestrator to coordinate them.

Benefits claimed:

1. Bugs/localization: failures are easier to diagnose in smaller components.
2. Improvements accumulate: updating a shared sub-skill improves all workflows that use it.
3. Don’t rebuild: create reusable building blocks once and reuse across projects.

Example mentioned: Claude previously re-wrote the same Python script for slide styles each session. Saving that script as a tool inside the relevant skill prevents repetition and token waste.

Central rule in this section: If a task can be implemented with code, implement it with tools/code so the system runs it reliably without regenerating it through tokens each time.

Rule 4: Skills get smarter every session

• Because skills persist, improvements compound over time.
• After a result is produced, the workflow is:
  • Decide whether the change is one-time or should become permanent in the skill.
  • Update the skill immediately (described as ~30 seconds of work).

The video describes asking Claude:

• “What exactly can I take into this skill?”

Then applying the suggested edits so future sessions perform better.

Example: replying to a client email

• Without a skill: you must restate tone, constraints, phrasing, and context every time (token-heavy).
• With a skill: those instructions are embedded once; later replies require fewer tokens.

Deliverables / Tutorial Elements Mentioned

• The creator promised:
  • A template for Claude Code instructions
  • The necessary skills and libraries that apply all four rules.
• The template/presentation is suggested to be available via a pinned message on the creator’s Telegram (link in description).

Main Speakers / Sources (Implied)

• Anthropic engineers (not individually named in the subtitles; referenced via quoted comments)
• The video narrator/creator (the person who introduces the four rules, provides the template, and directs viewers to Telegram)