AI WORKFLOWS

Claude Agent Skill - Session Context

Three markdown files. Zero re-explaining. Pick up exactly where you left off.

CLASSIFICATION:AIToolingClaude CodeProductivity
§00
ABSTRACT

Context loss between Claude Code sessions is annoying. You close a conversation, come back later, and spend 10 minutes explaining where you were. I fixed it with three markdown files that Claude reads automatically. Now I just pick up where I left off.

§01
EQUIPMENT & MATERIALS
TECH STACK:
[01] Claude Code[02] Markdown[03] Claude Skills
§02
THE PROBLEM

Every Claude Code session starts fresh. That's fine for quick tasks. For anything that spans multiple sessions? Pain.

You end up explaining the project again. What you were building. What decisions you made. Where you stopped. It's a tax on every conversation.

I wanted to open a new session and have Claude already know the state of things.

§03
THE THREE-FILE SYSTEM

Three files, three jobs:

CLAUDE.md gets read automatically at session start. Project conventions, dev commands, and instructions to check the other two files.

overview.md is the static plan. Goals, modules, architecture. Only changes when scope changes.

context.md is the live state. Current status, what just happened, what's next. Gets updated after every file operation.

§04
KEY DESIGN DECISIONS
  • 01Attempted vs Completed tracks both. If a session dies mid-task, the next one knows exactly where it broke.
  • 02Archive System keeps the active list under 20 items. Done work moves to an archive. Prevents the context file from becoming a mess.
  • 03Paused/Parked separates "stopped" from "finished". Blocked work gets its own section.
  • 04Verify on Start makes Claude summarize and ask "Does this match?" before doing anything. Catches stale state.
§05
HOW IT WORKS

Session starts. Claude reads CLAUDE.md automatically, then checks context.md. It summarizes: "We're on module 3, last thing completed was the API endpoint, next up is tests. That right?"

I confirm. We keep moving.

Before any file change, Claude logs what it's about to do. After, it logs what it did. End of session, the RESUME HERE section is accurate.

§06
RESULTS

The "context tax" is gone. I can close a conversation, come back a week later, and Claude picks up mid-task.

The mechanical triggers mean I don't maintain the files manually. Claude does it as part of normal work. That's the whole point. Systems that run themselves.