Context Engineering Kit stores project state in files. AI reads them every session. No more cold starts.
You: This project is a VAT system using Tauri... You: We use Zhipu API for OCR and translation... You: Last week we decided against SQLite... You: There's a bug in PDF parsing... You: We're working on batch recognition... AI: Got it. What do you need? You: (finally start working, 15 min later)
AI: Read CLAUDE.md, DECISIONS.md, memory/... AI: AI: Phase: MVP development AI: Last: batch invoice recognition (30 min ago) AI: Bug: PDF parsing garbled on some files AI: Next: fix PDF issue, then add progress bar AI: AI: Ready. What should I work on?
The real cost of AI amnesia isn't minutes. It's wrong decisions, repeated bugs, and broken continuity.
Every session picks up where the last one ended. Progress, blockers, next steps — all preserved.
Before: "What was I doing last Friday?" After: /start reads memory/current_state.md Last: auth refactor (2 days ago) Next: write tests for login flow
Recorded decisions don't get relitigated. AI reads why you rejected a proposal and never suggests it again.
Before: AI: "Let's use MongoDB!" You: "We discussed this last week..." After: You: "Check DECISIONS.md D001" AI: "Right, rejected MongoDB for SQLite."
Known bugs stay visible. AI won't reintroduce a fix or ignore a known issue when writing new code.
Before: AI fixes bug, next session re-introduces the same broken pattern After: /start shows active bugs AI avoids known-broken patterns /wrap moves fixed bugs to resolved
The kit installs project docs, memory files, prompts, and Claude Code commands. The files hold context; the commands read and maintain it.
Claude Code reads CLAUDE.md. Codex and other repo-instruction tools can read AGENTS.md. Both point AI at the same project context.
Architecture, decisions, and current tasks live in markdown. AI can see what exists, why choices were made, and what should happen next.
Progress, bugs, daily logs, lessons learned, and coding rules stay outside chat history. /start reads them; /wrap updates them.
If any of these sound familiar, yes.
Every minute counts. You can't afford to spend 15 minutes re-explaining the project to AI each session. /start gives it all back in 30 seconds.
Team mode shows who's working on what. Personal memory is gitignored, shared decisions use merge=union. No merge conflicts on state files.
Context doesn't expire. Open the project, run /start, and AI knows where you left off — even if it was weeks ago.
AI remembers which tutorial step you're on, what worked, what didn't. Your learning progress becomes part of the project context.
.claude/memory/) stores user preferences and general background — like "I prefer TypeScript" or "use conventional commits". This kit stores project state: what's done, what's broken, what was decided and why, what's next. They complement each other. Built-in memory is about who you are; this kit is about where the project is.CLAUDE.md, AGENTS.md, DECISIONS.md, TASKS.md, and ARCHITECTURE.md are committed — they're useful documentation for humans too. In team mode, personal memory/ files are gitignored. Everything is plain markdown, no binary or generated code.AGENTS.md gives Codex a repo-level entry point. Cursor and Windsurf can use the same facts through their rules systems; see the cross-tool section below..claude/commands/ are plain markdown too — change the steps to match your workflow.install.sh --team --user yourname. It creates shared memory, personal memory, memory/TEAM.md, and git merge rules. Existing files are not deleted; move useful solo memory into the team structure when you establish the team baseline./init-context analyzes your existing codebase and generates context files. If you already have CLAUDE.md, it reads and preserves it, only adding missing sections.# In your project root
git clone https://github.com/Vincent-chao-lang/context-engineering-kit /tmp/cek
/tmp/cek/install.sh
# In Claude Code /init-context # AI analyzes your project, generates CLAUDE.md, AGENTS.md, ARCHITECTURE.md, TASKS.md, DECISIONS.md, memory/... # Review and commit git add . && git commit
# Start of day /start # AI reads context, ready to work # ... code, commit, code ... # End of day /wrap # AI updates memory for next time
Same files. Different entry points.
Books and demo material for learning, sharing, and launching the methodology.