Doc
The 1000xagent master agent setup
How the 1000xagent master Claude Code agent is structured — one global agent that owns ~/.claude/ and orchestrates every project-level agent in ~/dev/. The live blueprint any engineer can clone for their own master agent.
The 1000xagent master agent is one global Claude Code agent that owns ~/.claude/
and orchestrates every project-level agent under ~/dev/. This is the live blueprint —
the directory tree, the agent roster, and the setup wizard are generated from the real
folder, so the doc mirrors the actual system.
Set up your master agent
What do you have today?
Mode 1: Setup from scratch — no global agent, no project-level agents
Step-by-step
11 steps. Each step has a prompt you paste into your own Claude Code session, plus any markdown spec files the prompt references. Save downloads to ~/.claude/specs/downloads/.
Step 1
Confirm your goal: a 1000x master agent
You are helping me set up my global Claude Code agent following the 1000xagent pattern. My goal: I want a single master agent at ~/.claude/ that: 1. Knows what I'm building across every repository in ~/dev/ 2. Can create new project-level agents on demand, each scoped to one repo 3. Acts decisively (autonomous within scope) but pauses for genuine creative direction 4. Self-enhances over time by capturing my preferences and corrections 5. Operates uniformly across software, business, content, and life-ops work Before we start, confirm you understand this goal in your own words, then ask me ONE clarifying question about my workflow that would change how we set this up. Just one. Then wait for my answer before proceeding.
Step 2
Enable bypass-permissions (4-layer setup)
Set up bypass-permissions across all four defense layers documented in the attached permission-policy.md spec: 1. Shell alias in ~/.bashrc: alias claude='claude --permission-mode auto --dangerously-skip-permissions' 2. Global ~/.claude/settings.json with: permissions.defaultMode: "bypassPermissions" permissions.allow: ["Bash","Read","Edit","Write","Glob","Grep","WebFetch","WebSearch","Agent"] skipDangerousModePermissionPrompt: true (top-level) 3. PermissionRequest auto-approve hook in settings.json hooks.PermissionRequest 4. Per-project .claude/settings.json (canonical template) in every repo under ~/dev/ Verify each layer is in place. Show me the diffs you intend to make and wait for approval before writing anything.
Files this prompt references
Step 3
Author ~/.claude/CLAUDE.md (your master agent identity)
Create my ~/.claude/CLAUDE.md based on the attached IDENTITY.md and CLAUDE.md reference. Adapt to my situation: - My handle/agent name (ask me if unsure) - The businesses I'm running or planning (ask me; don't make up names) - My operating principles (use the 1000xagent reference as a starting point, flag anything I should diverge on) - Folder structure section can be largely copied — that's the canonical layout Show me the draft. I'll edit voice and specifics before we save.
Files this prompt references
Step 4
Install the 5-piece project-agent contract
Save the attached project-level-agents.md rule to ~/.claude/rules/project-level-agents.md. This rule defines the 5 pieces every repo under ~/dev/ needs: 1. {repo}/.claude/CLAUDE.md + .claude/commands/{agent-name}.md 2. {repo}/.claude/rules/{at least one rule file} 3. Registry entry in ~/.claude/1000xagent/agents.yaml under agents: 4. Row in master ~/.claude/CLAUDE.md agent table 5. {repo}/.claude/settings.json with bypass-permissions template Going forward, when I ask you to create a new project, you produce all 5 pieces atomically. Confirm you understand by listing the 5 pieces back to me.Files this prompt references
Step 5
Install the autonomous-authority rule
Save the attached autonomous-authority.md to ~/.claude/rules/autonomous-authority.md and single-option-default.md to ~/.claude/rules/single-option-default.md. These define when you act vs when you pause: ACT: routine implementation, code changes, refactors, file edits, scaffolding, testing, deploys, anything reversible inside our scope. PAUSE for these four triggers ONLY: 1. Real creative/design direction (my voice, brand, copy) 2. Genuine ambiguity in the request (multiple plausible directions, hard to reverse) 3. Novel/first-of-its-kind work (after first execution, ship autonomously) 4. External constraint you can't resolve (password, browser-only step, physical action) For everything else: pick the best path, ship, let me interrupt if wrong. Never present a 3-option menu — pick one, justify in one sentence, execute. Confirm you have these rules loaded.
Files this prompt references
Step 6
Initialize ~/.claude/1000xagent/agents.yaml
Create my ~/.claude/1000xagent/agents.yaml using the attached agents.yaml as the schema reference. Start it with just my master agent block (matching my handle from step 3) and an empty businesses block. Schema requirements: - businesses: map with at least one key per business (each has name, type, agents list, status, public flag) - agents: map with one entry per agent (each has name, type, category, business, repository, remote, description, invocation, status, site block) For my master agent's category use "platform". For business use "{my-handle}-platform" pattern. Show me the file before saving. Don't invent agents I haven't asked for yet.Files this prompt references
Step 7
Author ~/.claude/1000xagent/KNOWLEDGE.md and folder-map.md
Create my ~/.claude/1000xagent/KNOWLEDGE.md (system architecture overview — what's in my ~/.claude/, what conventions I follow, what tools I use) and ~/.claude/1000xagent/folder-map.md (structural map of my .claude folder). Use the attached references as the structural template. The reference content is Tyler-specific — replace with my actual stack and structure. If you don't know something about my setup, ASK me one specific question rather than guessing.
Files this prompt references
Step 8
Save SKILL-TEMPLATE.md (the canonical skill scaffold)
Save the attached SKILL-TEMPLATE.md to ~/.claude/skills/SKILL-TEMPLATE.md. This is the canonical 2026-style skill scaffold: frontmatter + Phase 1 context-load + status table + pre-write gate. Going forward, when I ask you to create a new skill, follow this template — particularly the Phase 1 context-load step (always read source files BEFORE asking questions). Also save the attached skill-management.md and slash-command-tool-sequencing.md rules to ~/.claude/rules/ so the authoring discipline is auto-loaded.
Files this prompt references
Step 9
Carve out ~/.claude/specs/ namespace
Run: mkdir -p ~/.claude/specs/downloads Then save the attached specs-folder-conventions.md to ~/.claude/rules/specs-folder-conventions.md. This codifies the ~/.claude/specs/ partition: inbox/outbox (reserved for the future Maildir spec-exchange protocol between Claude Code sessions), downloads (where you save markdown files downloaded from this site), root (reserved). This is the directory the prompts on this site reference when they tell you to save a download.
Files this prompt references
Step 10
Install auto-commit + auto-push policy
Save the attached auto-commit-and-push-policy.md to ~/.claude/rules/auto-commit-and-push-policy.md. Adapt the per-repo auto-push table to my repos (you have none registered yet, so the table starts with just ~/.claude/ → auto-push YES). When I add project repos later, add a row each time based on whether the repo deploys to production on push (NO auto-push) or not (YES auto-push).
Files this prompt references
Step 11
Init the ~/.claude/ git repo + first commit
Initialize ~/.claude/ as a git repo: cd ~/.claude git init echo 'private/' > .gitignore echo 'projects/' >> .gitignore echo 'state/' >> .gitignore echo 'logs/' >> .gitignore echo '.credentials.json' >> .gitignore echo 'history.jsonl' >> .gitignore echo 'debug/' >> .gitignore echo 'file-history/' >> .gitignore echo 'plugins/' >> .gitignore echo 'ide/' >> .gitignore git add -A git status --short Show me the staged files before committing. If anything looks like it shouldn't be committed (secrets, machine-specific paths), flag it. Once clean, commit with: "init: 1000xagent master agent setup". Don't push yet — no remote configured.
Live directory tree of ~/.claude/
A filtered view of the master agent's folder — private paths (credentials, logs, session state) are excluded. Hover any file for its frontmatter description.
0
0 commentsNo comments yet.