claude-context-starter

Most people paste their context into Claude's project instructions and forget about it. It goes stale. They lose track of what's there. When something changes — a new job, a new project, a new workflow — updating it means hunting through UI fields with no version history. And if you use Claude in multiple places, you're maintaining the same context in multiple places.

This repo flips that. Your context lives in files. Claude Code reads them directly; claude.ai projects read the same files as uploaded knowledge. Update a file once, and every interface you use stays current. Build a skill once — like the avoid-ai-writing skill included here — and it works as a slash command in Claude Code and as uploaded context in any claude.ai project.

It also gives Claude a persistent, file-based auto-memory that grows over time (typed entries for user/feedback/project/reference) plus a /dream curator that runs autonomous passes over the memory dir — rot detection, contradiction surfacing, pattern capture — and produces reviewable proposals before anything writes back. See docs/auto-memory.md and docs/dream-architecture.md.

Quick start

git clone https://github.com/conorbronsdon/claude-context-starter.git my-context
cd my-context
bash scripts/setup.sh

The setup script walks you through the basics: sets your name, swaps the git remote to your own repo, installs hooks, and offers to launch Claude Code. If you don't have Claude Code yet:

npm install -g @anthropic-ai/claude-code

Once you're in Claude Code, run:

/setup

Claude interviews you and builds your context files — identity, first project, weekly priorities. Takes about 10 minutes. No manual editing required.

Using claude.ai instead? Open SETUP-PROMPTS.md and paste the prompts there — same questions, same results, you just copy the output into files manually.
Note: you can give your claude.ai session this repo as project context after it's set up, and easily sync it

After setup

Run your first session:

cd /path/to/my-context
claude
/start

/start loads your state and gives you a briefing. At the end of your session, run /end to log what happened. That's the core loop — /start → work → /end.

What's in the repo

CLAUDE.md                         # Root context — loaded on every session
ROUTING.md                        # Context routing for tasks without a slash command
TODO.md                           # Task backlog
SETUP-PROMPTS.md                  # Setup prompts for claude.ai users
identity/                         # Your bio, background, goals
projects/                         # Project context and skills (with worked example)
writing/skills/                   # Writing skills (avoid-ai-writing included)
state/                            # Session state, priorities, decisions, blockers
sessions/                         # Per-day session logs (created by /end)
inbox/                            # Drop zone for raw notes (triaged by /capture)
content/log.md                    # Published content log
commands/                         # Slash command definitions (including /setup, /dream, /skill-creator)
skills/                           # Cross-cutting / meta skills (e.g., skill-creator)
scripts/                          # Setup, validation, repo map generation
scripts/dream/                    # Curator prompts + how-to for the /dream substrate
docs/                             # Architecture guides — auto-memory, dream, migration, safety
references/                       # Integration setup (Google Workspace, Notion)
.claude/hooks/                    # Session start + SSOT guard + parallel-session guards
.github/                          # CI validation + PR template

# Auto-memory lives outside the repo (per-machine, often confidential):
~/.claude/projects/<encoded-cwd>/memory/
  MEMORY.md                       # Index loaded into every conversation
  <topic>.md                      # Detail files loaded on demand
  .dreams/<ISO>/                  # Curator proposal artifacts

Skills work everywhere

A skill is a markdown file that tells Claude how to do a recurring task. Build it once and it works as a slash command in Claude Code and as uploaded knowledge in claude.ai.

The avoid-ai-writing skill is included as a working example. In Claude Code: /clean-ai-writing. In claude.ai: upload writing/skills/avoid-ai-writing/SKILL.md as project knowledge and ask Claude to apply it.

To build your own:

1. Run /skill-creator and describe what the skill should do — it generates the SKILL.md, the command file, and the CLAUDE.md additions for you to review.
2. Or do it manually: read docs/agent-template.md, create projects/<your-project>/skills/<your-skill-name>/SKILL.md, add a command file in commands/, add a row to CLAUDE.md, then run scripts/validate-skills.sh.

See projects/README.md for conventions and the example musician project for the full pattern.

Auto-memory

Claude Code auto-loads ~/.claude/projects/<encoded-cwd>/memory/MEMORY.md at the start of every conversation in this project. The starter ships a spec (docs/auto-memory.md) for what to save (and what NOT to save) across four typed memory categories:

• user — role, expertise, preferences
• feedback — guidance about how to work, both corrections and validated approaches
• project — in-flight work, decisions, the why behind them
• reference — pointers to external systems (trackers, dashboards, channels)

MEMORY.md is an index. Detail files load on demand. Cap the index at ~100 lines.

docs/memory-template.md is the seed file. Copy it to the memory dir on first setup.

/dream — autonomous curator

Memory accumulates faster than humans review it. /dream runs an autonomous curator pass over the memory dir (default curator: rot detection — flags project memories that no longer match your state files or recent commits) and writes a reviewable proposal artifact. /dream-apply walks the artifact and applies accepted items, all under git on the memory dir so any pass is one git revert away.

See docs/dream-architecture.md for the full design (curator catalog, proposal schema, scope guards).

Running multiple Claude sessions

Running more than one Claude Code session against the same repo simultaneously will eventually corrupt your tree — silent branch-switches, one session's git add landing in another session's commit, files committed to the wrong branch. The fix is git worktree: one checkout per session.

The starter ships two hooks that enforce this pattern:

• worktree-guard.sh (PreToolUse) — blocks Edit/Write to a guarded repo's primary checkout when ≥2 Claude sessions are running. Worktrees are still free. Emergency override: touch .allow-shared-edit at the repo root.
• branch-hygiene.sh (SessionStart) — surfaces non-default HEAD on guarded repos so a silent branch-switch is noticed before any edits land.

Both hooks no-op until you list a repo basename in .claude/hooks/guarded-repos.txt. See .claude/hooks/README.md for setup.

Optional integrations

Google Workspace MCP — lets Claude read your calendar, email, Drive, and Sheets mid-session:

npm install -g @googleworkspace/cli
gws auth setup

The .mcp.json is already configured. See references/gws-mcp-setup.md for details.

claude.ai sync — upload CLAUDE.md, ROUTING.md, and relevant skill files to claude.ai projects as knowledge. See docs/claude-projects-sync.md for the workflow.

Validation

bash scripts/validate-skills.sh

Checks for: missing frontmatter, CLAUDE.md over 100 lines, committed secrets, stale files (90+ days).

Key conventions

• Single source of truth: Each fact lives in one file. Others reference, never duplicate.
• CLAUDE.md stays small: Under 100 lines. Detail goes in skills and ROUTING.md.
• Staleness dates: **Last Updated:** near the top of context files. Validation flags 90+ days.
• TODO.md vs. current.md: TODO is the full backlog; state/current.md is the top-of-mind view.

Migrating from existing Claude projects

See docs/migration-guide.md — includes an audit prompt, evaluation criteria, and restructuring guide.

Contributing

This is a template — the most useful contributions are structural: better examples, cleaner conventions, and skills others can adapt. Open an issue if you have a pattern worth adding.

Disclaimer

All views, opinions, and statements expressed on this account are solely my own and are made in my personal capacity. They do not reflect, and should not be construed as reflecting, the views, positions, or policies of Modular. This account is not affiliated with, authorized by, or endorsed by Modular in any way.