- Claude Code CLI installed
- A git repository you want to work in
Use agentic-workflow directly as your working directory.
git clone <agentic-workflow-repo-url>
cd agentic-workflowEdit context/USER.md with your details — name, role, preferences, tech stack.
Review context/SOUL.md. It defines how the agent behaves. Adjust to match your style.
claudeThe heartbeat runs automatically when CLAUDE.md is read. It loads your context, creates today's daily memory file, and greets you.
Say "wrap up" or run /meta-wrap-up. The agent reviews what was done, updates memory, and distills learnings.
Use /meta-bootstrap to install the workflow into any project, connecting it to a shared documentation repo.
~/repos/
├── agentic-workflow/ # Clone this repo
├── my-project/ # Your existing project
└── my-docs/ # Documentation repo (existing or new)
Start Claude Code in the agentic-workflow/ directory, then run:
/meta-bootstrap ~/repos/my-project --docs-repo ~/repos/my-docs
The bootstrap will:
- Create
workflow.jsonin your project (links to docs repo) - Create
context/directory with SOUL.md, USER.md template, MEMORY.md - Copy templates to the docs repo
- Install all workflow skills into
.claude/skills/ - Add the heartbeat section to your project's CLAUDE.md
- Update
.gitignorefor personal context files
The bootstrap adds these entries to .gitignore — verify they're present:
# Agent workflow — personal context (don't commit)
context/USER.md
context/MEMORY.md
context/memory/context/SOUL.md is intentionally not gitignored — it's shareable project config, not personal data.
cd ~/repos/my-project
# Edit context/USER.mdcd ~/repos/my-project
claudeThe heartbeat runs, loads context, reads workflow.json, and you're ready.
If your project already has a large codebase but little documentation:
/doc-init
This analyzes your code, git history, dependencies, and domain model, then generates:
- C4 architecture diagrams (system context, containers, components)
- Project evolution timeline (phases, decisions, dependency changes)
- Inferred ADRs from architectural choices found in code
- Domain model reference
- API endpoint reference
- Developer onboarding guide
All generated docs are marked as drafts for review. See Doc Init for details.
When you open Claude Code in a bootstrapped project, the heartbeat runs automatically:
1. Read context/SOUL.md → Agent identity and behavior rules
2. Read context/USER.md → Your profile and preferences
3. Read context/MEMORY.md → Long-term curated knowledge
4. Read today's daily memory → What happened earlier today
5. Read yesterday's daily memory → What happened yesterday
6. Read workflow.json → Where the docs repo lives
7. Check freshness → Flag stale files (>30 days old)
8. Greet you → Brief greeting with relevant context
If the heartbeat didn't run (e.g., first time, CLAUDE.md not yet read), invoke it manually:
/meta-heartbeat
Say "wrap up" or run /meta-wrap-up:
1. Review deliverables → git status, changed files
2. Update daily memory → Goal, work done, decisions, learnings, open items
3. Documentation check → Suggest ADRs for decisions, C4 updates for architecture changes
4. Distill to long-term memory → Move important insights to MEMORY.md
5. Collect feedback → "Anything to do differently next time?"
6. Commit (if requested) → Clean commit message, no AI branding
7. Summary → What was done, what's left
Use the doc-* skills to generate documents from templates:
/doc-prd "User authentication feature" → Product Requirements Document
/doc-rfc "Migrate from REST to gRPC" → Request for Comments
/doc-design "Payment processing system" → Design Document
/doc-adr "Use PostgreSQL over CosmosDB" → Architecture Decision Record
/viz-c4-diagram "Payment system" container → C4 Container Diagram
/doc-runbook "Database failover" → Operational Runbook
/doc-postmortem "2026-03-15 API outage" → Blameless Post-Mortem
/doc-spike "Evaluate message queue options" → Spike Report
Each skill reads workflow.json to determine where templates live and where to output the generated document. If workflow.json doesn't exist, it falls back to local templates/ and docs/ directories.
After pulling a new version of agentic-workflow:
cd ~/repos/agentic-workflow && git pullThen in any bootstrapped project:
/meta-upgrade
This compares your installed skills against the source repo, shows what changed (new skills, updates), and lets you selectively apply upgrades.
After bootstrapping, your project looks like:
my-project/
├── CLAUDE.md # Your existing CLAUDE.md + heartbeat section
├── workflow.json # Points to docs repo and agentic-workflow
├── context/
│ ├── SOUL.md # Agent identity (shareable)
│ ├── USER.md # Your profile (gitignored)
│ ├── MEMORY.md # Long-term knowledge (gitignored)
│ └── memory/ # Daily session logs (gitignored)
│ ├── 2026-03-16.md
│ └── archive/
├── .claude/
│ └── skills/
│ ├── meta-heartbeat/ # Session start
│ ├── meta-wrap-up/ # Session end
│ ├── meta-upgrade/ # Sync from agentic-workflow
│ ├── doc-prd/ # Document generation
│ ├── doc-adr/ # ...
│ ├── viz-c4-diagram/ # ...
│ └── your-custom-skill/ # Project-specific skills (untouched by upgrade)
└── src/ # Your actual code
- Architecture — understand the three-layer context loading
- Context Files — learn what goes in SOUL.md, USER.md, MEMORY.md
- Skills System — understand skill structure and naming
- Multi-Repo Setup — connect multiple projects to one docs repo
- Daily Workflow — day-to-day usage tips