Autocontext

A Claude Code plugin that gives your AI assistant a memory for each project.

The problem

Every time you start a new Claude Code session, Claude starts from scratch. It doesn't remember that the database migration broke last time you used mocks, that the deploy script needs a cache clear first, or that the API returns dates in a weird format. You end up correcting the same mistakes across sessions, and those corrections vanish when the conversation ends.

What autocontext does

Autocontext listens to your conversations and captures the moments where you correct Claude, "no, use the other API", "you forgot to clear the cache", "that's the wrong import path." It saves those corrections as structured lessons in a JSON file inside your project.

The next time you open a session in that project, autocontext loads the relevant lessons and uses them to avoid repeating the same mistakes. Lessons that keep proving useful gain confidence over time. Lessons that go stale lose confidence and eventually fade out.

It's like giving Claude a notebook that it reviews before starting work each day.

What it looks like in practice

• You tell Claude "no, always use pytest -x here, not pytest" → autocontext captures that as a lesson candidate
• Next session, when Claude is about to run pytest, autocontext injects a warning: "lesson: always use pytest -x in this project"
• If Claude follows the lesson without you correcting it again, the lesson's confidence goes up
• If you correct it differently, the old lesson decays

Over weeks of use, your project builds up a curated knowledge base of project-specific patterns, gotchas, and preferences that Claude loads automatically.

Install

# Clone claude-skills-journalism anywhere, then install the autocontext
# skill into ~/.claude/skills/ (Claude Code discovers skills one level deep):
git clone https://github.com/jamditis/claude-skills-journalism ~/projects/claude-skills-journalism
mkdir -p ~/.claude/skills
cp -r ~/projects/claude-skills-journalism/autocontext ~/.claude/skills/
# or: ln -sfn ~/projects/claude-skills-journalism/autocontext ~/.claude/skills/autocontext

Claude Code loads the skill automatically on next launch.

Quick start

1. Run /autocontext:setup to configure your preferences (identity, sensitivity, loading behavior, 12 steps total).
2. In any project, run /autocontext:init to create the .autocontext/ directory and start accumulating lessons.
3. Just use Claude Code normally. Autocontext runs in the background.

Commands

How it works under the hood

Autocontext uses Claude Code's hook system to run at five points in every session:

1. When a session starts, loads lessons from .autocontext/lessons.json, ranked by relevance to the files you've been working on. If there are lesson candidates from your last session, a curator pass decides which ones are worth keeping.
2. Before each edit or command, checks if any loaded lessons are relevant to the file being edited or the command being run. If so, it injects a short warning so Claude doesn't repeat a known mistake.
3. When you type a message, pattern-matches for correction phrases like "no, use X instead" or "you forgot to..." and queues them as lesson candidates.
4. After test file edits, runs quality checks on test code (catches tautological assertions, happy-path-only test suites, and other common test smells). Also tracks build/test times against baselines to flag performance regressions.
5. When a session ends, lessons that were loaded and not contradicted get a confidence boost. The playbook summary file is regenerated.