Skip to content

Claude Code

Updates

03/17/2026

An agentic coding assistant which not only addresses coding, but also discovery and design.

The architecture of the assistant looks like:

  • CLAUDE.md file represents memory across sessions. It can include style guidelines and common commands. This file is automatically loaded into the context.
  • See overview for installation and getting started.
  • Claude Code offers a variety of settings to configure See the settings article. See also ~/.claude/settings.json

Capabilities

  • Conversation history is stored locally, which is loaded into context.
  • Can launch sub-agents
  • Manage to dos
  • Refer back to previous information
  • Agentic search
  • A marketplace is a collection of plugins that can be installed into Claude Code. Plugins extend Claude Code's capabilities with company-specific knowledge, tools, and workflows.
  • Plugins can contain skills, hooks, agents and commands.

Integrating with existing git repo.

Here are 3-step playbook for getting Claude Code productive and safe on existing repos.

  • Create a first CLAUDE.md file. Don't use /init as-is — it produces bloated output. As good practices:
Include Exclude
Shell commands LLM can't guess Anything LLM can figure out from code
Code style rules that differ from defaults Standard language conventions
Testing instructions and preferred runners Detailed API docs (links to existing content instead)
Architectural decisions specific to your project Long explanations or tutorials
Common gotchas or non-obvious behaviors File-by-file codebase descriptions
  • Adopt TDD
  • Plan First, then Scale with skills and plugins.
    • Planning is most valuable when you're uncertain about the approach or the change spans multiple files. Ctrl-G to update the plan.
    • Prefer CLIs over MCP
    • Leverage plugin like Superpowers
    • Manage context aggressively — Use /clear between unrelated tasks, /compact to summarize mid-session, and subagents for investigation

Useful prompts

Goal Prompt
Fix build issue The build fails with this error: [paste error]. Fix it and verify the build succeeds. Address the root cause, don't suppress the error.
Fix runtime issue Use subagents to investigate why [service] is throwing [error]. Check logs and recent changes.
Code review Read ... folder, What are the key files and flows?
Build new feature I want to build [brief description]. Interview me about edge cases, tradeoffs, and implementation using AskUserQuestion. Then write a spec to SPEC.md
Review PR (collaborate on PR) Review the PR at [link]. Look for edge cases, race conditions, and consistency with existing patterns
Address PR feedback Read the review comments on PR #[number]. Address each one, reply to the threads, and push fixes.

Commands

commands Role
/init analyze a repo and create CLAUDE.md
/clear At ~40% context usage, expect small inconsistencies. At ~70%, expect bigger mistakes. Use /clear between unrelated tasks
/models Select a model

Running Claude with Google Vertex/AI

Be connected to VPN (GlobalProtect or Twingate) 

gcloud auth application-default login
claude
claude  --model=claude-haiku-4-5
# Prefer opus for coding

See GCP pricing plan.

Running with Local LLM

  • Create a file named Modelfile 
    echo "FROM gemma4:26b
PARAMETER num_ctx 65536" > Modelfile
  • Create the optimized version 
    ollama create gemma4-coding -f Modelfile
  • Start claude: 
    ANTHROPIC_BASE_URL="http://localhost:11434/v1" ANTHROPIC_API_KEY="local" claude --model gemma4-coding

This will be slow and will have some challenge with the plan model of Claude and gemma4.

Wiki LLM

See the core principles. See techno/claude/llm-wiki.md

Sources