Field notes on agentic workflows, context engineering, and the plumbing that makes long-running Claude Code sessions actually useful.
After ninety days of Claude Code, you have a hundred sessions you cannot find, a dozen projects you cannot resume, and an agent that forgot all of it. That is context debt. This is the post that names it, and the one command that pays it off.
Everything you need to get the most out of Claude Code, from project setup and session management to cost control, memory, MCP tools, and advanced patterns that compound over time.
The point of a memory layer isn't search. It's reflex transfer. Every steering decision you captured becomes the agent's starting position on the next problem.
You're not prompt engineering. You're steering. Here's what context engineering actually means when you're running twenty-plus apps and thirty terminals at once.
You do not have one project open. You have twenty. You do not run one Claude session. You run thirty. Here is the shape of the workflow that does not collapse under that load.
You have 1,200 sessions and climbing. Aliases, tags, collections, pins, notes. Here is the mental model that actually scales, with examples from a real power-user setup.
The feature nobody else has: `recall context <id> | claude` hands a full earlier conversation to a fresh Claude session as starting context. Here is the tactical guide.
Timing windows are how you lose data. Walking the process tree is how you do it deterministically. Here is the algorithm inside the Claude Recall VS Code extension (v0.7).
Offline. On-device. No servers. In 2026 it sounds nostalgic. Here is the case for why local-first is the only responsible posture for tools that read your code and your conversations.
MCP is Anthropic's open protocol for letting models call native tools. Here is what it does, why it matters, and a worked example using Claude Recall's six read-only tools plus the seven write tools we shipped in v0.13.
VS Code, Cursor, and Windsurf all cap integrated terminal scrollback around 1,000 lines. A real Claude Code session blows past that in ten minutes. Here is what actually breaks and how to work around it.
Claude Code writes every conversation to ~/.claude/projects/ as JSONL. Here is exactly what lives there, what each file contains, and why it is useless by default.