Heidelberg AICurriculum

CLAUDE CODE ADVANCED FULL COURSE (3 HOURS)

Nick Saraev ·2026-03-28 ·2 min read

Summary written by us from the video's transcript. The video, and everything in it, is Nick Saraev's work.

Learn how to structure and continuously improve Claude.md system prompts for advanced Claude Code projects, using global and local workflows to boost efficiency and reduce token costs.

Takeaways

  • Claude.md compresses workspace knowledge, preferences, capabilities, and failure logs into a single prompt file.
  • Use separate global and local Claude.md files: global for high‑level rules, local for project specifics.
  • Iteratively update the local Claude.md after each feature cycle to reduce token usage and speed up future work.
  • Periodically extract cross‑project patterns and add them to the global Claude.md after human review.
  • Keep system prompts concise, secure, and aligned with your preferred coding conventions.

What a Claude.md Is

Claude.md serves four core purposes: it compresses knowledge about your workspace so Claude can reason without scanning every file; it stores your personal preferences and conventions for how Claude should respond; it declares the capabilities Claude has within that project, preventing unnecessary “I don’t know” loops; and it logs past failures and successes to prune unproductive solution paths.

Global vs. Local Claude.md Files

Claude looks for two kinds of system‑prompt files: a global Claude.md in your user folder (injected on every session) and a local Claude.md inside a .claude folder of a specific repository (injected only when you work there). Use the global file for high‑level reasoning, personal preferences, and agency declarations; use the local file for project‑specific knowledge, API docs, and low‑level details.

Designing Effective System Prompts

Combine the four Claude.md principles into short, actionable bullet points. For example, in the global file you might specify “always return absolute file paths” or “use functional Rust style”; in the local file you could embed a compressed summary of the repo’s purpose and include the full GoHighLevel API spec to avoid extra tool calls. Keep prompts concise to conserve tokens while still providing enough context for Claude to act autonomously.

Local Workflow Loop

When starting a new feature, plan it, let Claude attempt implementation, then capture every success and failure. Condense those learnings into a high‑density update to the local Claude.md. Repeating this loop makes each subsequent iteration faster—often shaving 10 % or more of token usage as the model learns from its own history.

Global Workflow Loop

After many local cycles, extract patterns that appear across projects (e.g., recurring rabbit‑hole searches). Manually review these insights—because chaining multiple Claude steps reduces overall reliability—and add high‑ROI bullet points to the global Claude.md. This step ensures future projects inherit the distilled wisdom without re‑learning it each time.

Practical Setup in Antigravity

Install the Antigravity IDE from antigravity.google, add the Claude Code VS Code extension, and log in. The Claude logo will appear for quick agent launches. Both the desktop app and the extension respect the same global/local Claude.md files, so your workflow is consistent regardless of environment.

Maintaining Efficiency and Security

Regularly audit your global Claude.md to avoid bloating it with low‑value details; keep security‑sensitive tokens or credentials out of system prompts. When performance fluctuates, fall back to alternative agents or bundle multi‑agent orchestration as a safety net, but always retain a human review step for critical updates.