project-butler

project-butler

English | 中文

Make AI coding agents remember your project between sessions.

project-butler helps Claude Code, Cursor, Codex, and similar AI coding assistants behave like long-term project teammates instead of starting from scratch every session.

For normal use, you only need four actions:

/project-butler   Set up project memory
end session       Save progress and next steps
continue          Resume next time
status            Check where the project stands

For projects that need stronger product, architecture, roadmap, research, or eval alignment, project-butler can also create a Project Profile during setup and offer profile-aware Normal Close / Full Close behavior.

Quick Start

Install as a Claude Code skill:

git clone https://github.com/JamesShi96/project-butler.git ~/.claude/skills/project-butler

Open any project and set up project memory:

/project-butler

Work normally. At the end of a work session:

end session

Next time, resume without re-explaining the project:

continue

That is enough for daily use. For Cursor, Codex, and other assistants, see Tool Compatibility.

Updating project-butler

The skill auto-checks for updates on every invocation. Once per day per machine, it runs git fetch against its own repo and compares local HEAD to origin/main. If behind, Claude Code asks you once — and at most once every 24 hours:

project-butler is 3 commits behind upstream. Update now?
  › Update now       — pull the latest version right here
  › Remind me later  — ask again tomorrow
  › Stop reminding   — turn the check off

Picking Update now runs a fast-forward-only git pull for you and reports the result. If it fails (local changes, offline, SSH blocked), you get the manual command plus an HTTPS fallback. Nothing is ever pulled without you choosing it.

The prompt is worded in your project's CLAUDE.md Language: setting (English / Chinese / bilingual), and stops appearing once you are up to date — the cache is keyed on commit SHA, not just time.

Side effect on git status: after the auto-fetch, git status inside the skill directory may show "behind origin/main by N commits". This is expected and harmless — the skill never modifies the working tree unless you pick "Update now".

Silencing: export PROJECT_BUTLER_NO_UPDATE_CHECK=1. Only the literal value "1" silences — =0, =false, or empty does not silence (counter-intuitive but intentional).

Note: the check runs as a skill instruction, so it is best-effort — it may occasionally not fire. Run the script manually any time you want a definitive answer.

Debugging a missing prompt: from an external shell only, run the shared update-check script with PROJECT_BUTLER_UPDATE_CHECK_DEBUG=1. Never enable debug inside Claude Code — CC captures stderr into the LLM context and debug output will leak into responses.

Cursor / Codex: these tools do not have Claude Code's skill lifecycle, so update checks are manual/on-demand:

bash "${PROJECT_BUTLER_SKILL_DIR:-$HOME/.claude/skills/project-butler}/scripts/check-update.sh"

Outside Claude Code there is no prompt — the script just prints a VERSION_NOTICE: block with the update command when you are behind, and nothing when you are current.

Reach limitation: if you installed project-butler before v1.7.0, you do not have this auto-check feature yet. Pull once manually:

cd ~/.claude/skills/project-butler && git pull

After that, future updates are announced automatically.

Why It Exists

AI coding assistants are powerful in one session and forgetful across sessions. If any of these sound familiar, project-butler is for you:

• "I had to re-explain the architecture again." Each new session starts with missing context.
• "What did we decide about naming conventions last week?" Decisions disappear into chat history.
• "The README and TODOs keep drifting from reality." Project state stops matching the files.
• "The AI keeps violating rules I already explained." Rules live in your head instead of in project memory.
• "I switch between Claude Code, Cursor, and Codex." Different tools need one shared source of truth.

project-butler turns a project folder into that source of truth, so the next AI session can pick up where the last one stopped.

Main Commands

All triggers are natural language. Use slash commands only for first-time setup.

Command	Use it when
/project-butler	Set up or upgrade project memory.
end session / we're done	Save progress, refresh next steps, and record important changes.
continue / continue from last time	Resume the previous session without re-explaining context.
status / where are we	Get the current project state and the next best step.

Advanced Commands

Command	Use it when
continue full context	Rebuild the full project trajectory after a long break or assistant switch.
review claude / check the rules	Review candidate project rules before they become long-term rules.
sync wiki / update overview	Force-refresh PROJECT.md.
organize files	Clean up new files according to STRUCTURE.md.
change language	Switch project management files between English, Chinese, and bilingual mode.
normal close	Save the session and defer profile-impacting updates into the pending queue.
full close	Align affected profile docs now with a bounded Scope Plan.
profile setup / foundation repair	Create or repair the project profile and baseline reference docs after confirmation.

Session recovery (continue / continue full context) is routed through project-butler internally. There is no separate /continue command to install.

What It Maintains

Run /project-butler once. It maintains these plain Markdown files in your project:

project-root/
├── CLAUDE.md                   <- Project rules / constitution
├── PROJECT.md                  <- Current project wiki
├── STRUCTURE.md                <- File organization rules
├── UPDATE_LOG.md               <- Milestone-level changelog
├── DOCS.md                     <- Document index and metadata
├── session-handoff.md          <- Cross-session handoff
├── TODO.md                     <- Execution checklist
├── docs/                       <- Archived project documents
├── log/                        <- Session logs
└── .claude/
    ├── candidates.md           <- Candidate rules for review
    ├── project-profile.json    <- Project profile config
    ├── profile-pending.json    <- Profile pending/debt queue
    └── .file-snapshot.json     <- File organization snapshot

The core files are plain Markdown, so other tools can read them even when they do not run the skill natively.

What that means in practice:

• Keeps the current project state readable.
• Keeps next steps clear between sessions.
• Keeps project documents indexed and findable.
• Records milestone changes so the project has a clear history.
• Keeps new files from drifting into random folders.
• Preserves long-term rules only after user review.

Project Butler also keeps a small machine-readable profile so the assistant can understand which long-lived docs matter, which sections are protected, and which profile updates have been deferred.

Tool Support

Tool	Status	How it works
Claude Code	Native skill	Install this repo under ~/.claude/skills/project-butler and run /project-butler.
Cursor	Project rules, best-effort	project-butler can generate .cursor/rules/project-system.mdc, which points Cursor at the same project memory files and mirrors the main triggers.
Codex	AGENTS.md, best-effort	project-butler can generate AGENTS.md, which points Codex at the same project memory files and mirrors the main triggers.
Other AI assistants	File-based	Any assistant that can read project files can use the project memory as shared context.

See docs/compatibility.md for details and caveats.

How It Works

Internals: The Memory Stack

project-butler uses a 7-component memory stack internally, organized by stability:

Stable rules
┌─────────────────────────────────────┐
│  CLAUDE.md / project rules          │  <- Human-reviewed principles
│  ↑ candidates collected by AI       │
└─────────────────────────────────────┘
            ↑ distilled from work
Current state
┌─────────────────────────────────────┐
│  PROJECT.md                         │  <- What the project is now
│  STRUCTURE.md                       │  <- Where files belong
│  UPDATE_LOG.md                      │  <- Milestone-level changes
│  DOCS.md                            │  <- Document index
│  .claude/project-profile.json       │  <- Profile config
│  .claude/profile-pending.json       │  <- Profile debt queue
└─────────────────────────────────────┘
            ↑ summarized from facts
Raw facts
┌──────────────────────┐ ┌───────────────────────┐
│  log/                │ │  TODO.md              │
│  What happened       │ │  What needs doing     │
└──────────────────────┘ └───────────────────────┘
            ↓
session-handoff.md       <- Where the next session should resume

Bottom feeds top. Top constrains bottom.

• Session logs capture what happened.
• Handoff tells the next assistant where to resume.
• Project wiki summarizes the current state.
• TODOs keep execution visible.
• Rules / constitution preserve decisions that should keep guiding the project.
• Update log records significant changes at milestone level.
• Structure rules keep files from drifting into chaos.
• Document index keeps project documents organized under docs/.
• Project profile tracks project shape, document tiers, document policies, pending profile updates, and review queue items.

Project Profiles

Project Profile System is internal profile-aware behavior for setup and close. Fresh setup stays conversational and can remain lightweight by creating only minimal confirmed docs.

During setup, project-butler asks what you are trying to do in natural language, infers the project shape, asks a few targeted follow-up questions, and proposes Required / Recommended / Optional reference docs. It does not force you to pick a fixed project type or expose Profile System as a setup switch.

During close, profile-aware projects can use:

Mode	Behavior
Normal Close	Save the session and record profile-impacting changes in .claude/profile-pending.json.
Full Close	Read only affected profile docs, present a Scope Plan, and apply safe updates inside that boundary.

Full Close confirms boundaries, not every small edit. It still requires explicit confirmation before changing protected sections, document policies, stable baselines, or whole-document rewrites.

Language Support

project-butler supports three language modes:

Mode	Content language	User file naming
en	English	English naming (kebab-case)
zh	Chinese	Chinese naming allowed
bilingual	Chinese with English annotations	English preferred, Chinese acceptable

You choose the mode during setup, and can later say change language.