The Agentic Cookbook.

You are my Agent Orchestrator.
Interview me first, one question at a time.
Map my personal brand, business systems, files, apps, and recurring workflows.
Identify the agents I need based on the interview.
For each agent, define its identity, tools, memory, schedule, reporting, and success criteria.
After the interview, build the agent files, prompts, folders, commands, and setup steps for me.
Confirm before changing anything live.

Repeat this process for every agent I need built.

# <Department> Agent

## Mission
<One sentence — what this agent exists to do.>

## Identity
- Name:
- Reports to: Orchestrator
- Department:

## Owns
- <Outcome 1>
- <Outcome 2>

## Does Not Do
- <Hard boundary 1>
- <Hard boundary 2>

## KPIs
- <Measurable 1>
- <Measurable 2>

## Skills
- <skill-name>
- <skill-name>

## MCPs / Connectors
- <connector>
- <connector>

## Memory
- Context: memory/context/<dept>.md
- People:  memory/people/<dept>.md
- Project: memory/projects/<project>.md

## Schedule
- Triggers: <schedule | webhook | manual>
- Cadence:  <daily | weekly | on-event>

## Reporting
- Logs to: logs/<dept>/
- Reports to Orchestrator: <when + format>

## Definition of Done
- <Acceptance criteria>

Local MCPs:
  - NotebookLM
  - Playwright
  - Chrome DevTools
  - Vibe Prospecting
  - Auto-browser
  - Google Surf

Cloud Connectors:
  - Gmail
  - Google Calendar
  - Google Drive
  - Notion
  - Figma
  - Canva
  - Krisp
  - Mem
  - Pipeboard Meta Ads
  - Computer Use

Per-Department Mapping:
  Ads        : Pipeboard Meta Ads, Playwright, Chrome DevTools
  Analytics  : Google Drive, Notion, Krisp
  Comms      : Gmail, Google Calendar, Krisp, Notion, Mem
  Content    : Google Drive, Figma, Canva, NotebookLM
  Finance    : Google Drive, Notion
  Recruiting : Gmail, Google Calendar, Notion, Vibe Prospecting
  Sales      : Gmail, Google Calendar, Notion, Vibe Prospecting
  Social     : Playwright, Canva, Figma
  Logging    : none

Bring your own credentials. Never commit tokens.

/settings
  → Enable remote control and configure the working surface.

/ultraplan
  → Create an execution protocol before building.

/goal
  → Point the agent at the protocol and mission.

/agents
  → Build agents around the goal.

/ultrareview
  → Review output against the plan.

codex exec "<task + story path + test to satisfy>"
  → Run Codex as a build lane.

codex review (or a fresh Codex pass)
  → Run GPT-side review separate from build context.

WORKSPACE ROOT
│
├── CLAUDE.md ─────────────────────────  ◄── TIER 1 — HOT CACHE
│     • Lives at workspace root
│     • Auto-loaded at session start
│     • Kept SHORT: 50–80 lines max — it's an index, not a manual
│     • Acts as the hub — points down into Tier 2 when needed
│     • Defines project overview, tech stack, commands
│     • Contains coding conventions & architecture
│     • Supports CLAUDE.local.md overrides
│
├── Claude.local.md
│     • Local override of CLAUDE.md (gitignored, per-developer)
│
├── .mcp.json
│     • Stores MCP integration configs
│     • Connects to GitHub, JIRA, Slack, DBs
│     • Shared across team via git
│
├── .claude/   ◄── CLAUDE-FACING CONFIG (sits inside Tier 1)
│     │
│     ├── settings.json
│     │     • Controls permissions & tool access
│     │     • Defines model selection and hooks
│     │     • Supports settings.local.json overrides
│     │
│     ├── settings.local.json
│     │     • Local override of settings.json (gitignored)
│     │
│     ├── rules/
│     │     • Modular .md files by topic
│     │     • Covers style, testing, API design
│     │     • Can target specific files/paths
│     │     ├── code-style.md
│     │     ├── testing.md
│     │     └── api-conventions.md
│     │
│     ├── commands/
│     │     • Custom slash commands (/project:<name>)
│     │     • Used for repeatable workflows
│     │     • Supports shell execution
│     │     ├── review.md
│     │     └── fix-issue.md
│     │
│     ├── skills/
│     │     • Auto-triggered based on task context
│     │     • Loads only when needed
│     │     • Keeps context lightweight
│     │     └── deploy/
│     │           ├── skill.md
│     │           └── deploy-config.md
│     │
│     ├── agents/
│     │     • Specialized sub-agents with roles
│     │     • Isolated context windows
│     │     • Custom tools and model preferences
│     │     ├── codeReviewer.md
│     │     └── security-auditor.md
│     │
│     └── hooks/
│           • Event-driven scripts (pre/post tool use)
│           • Automates validation, linting, formatting
│           • Blocks unsafe operations
│           └── validate-bash.sh
│
├── memory/ ─────────────────────────  ◄── TIER 2 — DEEP MEMORY
│     • Lazy-loaded — only pulled when CLAUDE.md references it
│     • NEVER commit to git (sensitive + high churn)
│     • Use predictable slugs so the hub can point cleanly
│     ├── people/        (one .md per person)
│     ├── projects/      (one .md per project)
│     └── context/       (general context files)
│           └── credentials.md   (chmod 600)
│                 • Long-lived, revocable, scoped creds ONLY
│                   (e.g. WordPress Application Passwords)
│                 • Document the revocation path inline
│                 • REFUSE to store: real account logins,
│                   private keys, payment info
│
└── <client>-site-bundle/ ──────────  ◄── TIER 3 — WORKING ARTIFACT
      • Flat file mirror of the live site
      • Canonical source of truth — bundle wins over memory
      • Everything stored as .md / .json (no proprietary formats)