Example setup

An illustrative ecosystem, separate from your learning workspace.

Explore the example · Tech stack

Tools in this example

Tech stack

A tech stack is the set of tools used together. These cards explain what each AI tool is for, when it is useful, and its limits. To see how the tools connect, open the technology map. To see the jobs carried out using them, visit Bots & tasks.

xAI

Grok

Rung 2 — chat with connectors, plus Grok Build

Architect and builder of the playbook itself

Use it for
Designing loops, editing GitHub, drawing architecture, standing up a new automation.
In this practice
This dashboard. The how-i-do-it repo. Grant workflow files. Netlify deploy status. Trip Waffle itineraries over MCP. Architecture work in this chat; code that should live in git still goes through Cursor → GitHub.
Strength
Fast, connector-rich, willing to write the repo and the map in the same sitting.
Limits
A live chat is a poor alarm clock. Put cadence on the Bot, not in an open thread.
MCP (custom)Git pushDrive writeConnector / plugin

To copy. Use Grok to build the system. Use Grok Bot to run the system.

xAI · scheduled

Grok Bot

Rung 2 — persistent scheduled agent

Daily coach and secretary

Use it for
Anything that must happen Tuesday at 9, or when mail from a named domain arrives.
In this practice
Monday goals, weekday morning, midday, evening done list, Friday review, 90-day cascade, Inkbox secretary identity, Trip Waffle itineraries over MCP, Rex receipts review after trips.
Strength
Wakes on a clock, already holding MCP sockets, writes a dated trail.
Limits
A long prompt with too many tools becomes excessive agency.
MCP (custom)Event triggerDrive writeVoice loopConnector / plugin

To copy. One bot, many schedules, one instructions file in GitHub. Split tools by job if a job can send mail. Map them on the roster — not as one blob.

OpenAI

ChatGPT Pro

Rung 1–2 — Custom GPTs + direct connectors

Direct door into Google and GitHub, and a workflow door into Command Center

Use it for
Ad-hoc Drive/mail work when you are already in ChatGPT. Custom GPTs for a single role. Command Center workflow.
In this practice
ChatGPT Pro. Direct connectors as an alternate door — not the weekday cadence owner. Workflow into Command Center. Codex is the coding sibling, not a second chat.
Strength
First-party Gmail, Drive, GitHub, Calendar. Broad tool coverage, polished chat. Codex is the coding agent on the same subscription.
Limits
Vendor-locked. Easy to duplicate work the Bot already does.
Direct OAuthConnector / plugin

To copy. If the model offers a first-party Google login, draw a solid ‘direct’ line, not an MCP dash. Draw Codex as its own node.

OpenAI

Codex

Rung 2 — IDE agent (VS Code / Cursor extension)

The other coding resident. ChatGPT Pro's hands in the repo.

Use it for
Writing or reviewing anything that should become a git commit, when the OpenAI coding agent is the one on the job.
In this practice
Live. Native extension in VS Code or Cursor. One job, one agent — not beside Claude Code on the same change.
Strength
Lives in the editor. Same job as Claude Code: edit, run, commit.
Limits
A second coding agent on the same tree is how secrets and conflicting edits land.
Git pushConnector / plugin

To copy. Codex is hands. GitHub is memory. Draw it as its own node, the way Claude Code is drawn.

Anthropic

Claude

Rung 1–4 — Projects, Claude Code, plugins

Careful drafter, grant critic, instruction twin

Use it for
Grant critic, long playbook prose, syncing GitHub instructions into Drive, Claude Code for repos.
In this practice
salim-weekly-goals instruction twin. Grant pipeline (Drafter/Critic). Claude Code runs on the local drive in ~/projects: it reads the governance repo first, builds and ships this site, and commits with the causal chain. Wide plugin access, used narrowly.
Strength
Long documents, evaluation, subagents. Plugin tray covers Calendly, Dex, Todoist, Netlify, Inkbox, WhatsApp. Own MCP client for Wispr and Trip Waffle. Claude Code stitches in Cursor.
Limits
A plugin broker (Composio) is an extra trust hop — gap-fill only, never the default when a native plugin exists.
MCP (custom)Connector / pluginGit pushDrive write

To copy. Same GitHub instructions as Grok. Stamp Drive copies ‘synced from GitHub’. Claude Code works in the folder; the commit is still the handoff, and it does not push to main unless asked.

Perplexity

Perplexity

Rung 2 — Computer + Spaces

Cited research scout

Use it for
Grant Scout, trend watches, teaching-program source gathering, anything that must cite.
In this practice
Research briefs that should land in GitHub as Markdown. Teaching ladder demonstrations.
Strength
Live web with sources. Good orchestrator for research-only jobs.
Limits
Not the place for private mail or PHI. Not yet on the daily cadence.
On the horizon

To copy. Keep Perplexity on public knowledge. When a brief is done, commit it to the brain — do not leave it in a chat.

Google

Gemini

Rung 1 — Gems + Workspace

Native Workspace colleague

Use it for
When the file already lives in Drive and you want a Gem that only sees that folder.
In this practice
Partial. Workspace-native reads. Not the cadence owner.
Strength
Already sits next to Drive, Calendar, Gmail. Gems are the fastest Rung 1 agent.
Limits
Weaker at long autonomous coding than Claude. Easy to over-scope a Gem.
Direct OAuth

To copy. Give a Gem one folder, not My Drive. Draw those wires as direct, because they are.

Anysphere

Cursor

Rung 2 — IDE agent (Composer / Cursor Agent)

The workbench. Code is typed here, then pushed to the brain.

Use it for
Writing or reviewing anything that should become a git commit — the playbook site, grant scripts, HT Sherpa.
In this practice
How I Do It, grant-agent-workflow, HT Sherpa. Claude Code often stitches in the same window. GitHub remains the operative note; Cursor is not a second brain.
Strength
The model sits in the repo. Inline edits, terminal, and MCP servers in one window.
Limits
An agent with a shell is excessive agency if the tree contains secrets or prod credentials.
Git pushMCP (custom)

To copy. Cursor is hands. GitHub is memory. Push, then the other AI tools can read it. Draw a git line, not a chat line.

The difficulty ladder

Start on rung 1. Climb only when the previous rung genuinely does not fit. This playbook lives on rung 2 — scheduled bots and connectors — on purpose.

  1. 1Custom GPT / Claude Project / Gemini Gem / Grok BotEveryone15 minutes
  2. 2Perplexity Computer, Claude Code / Cursor Agent, Grok scheduled botsWhere this playbook lives30 minutes
  3. 3n8n / Make / Zapier / GumloopWhen the job crosses many SaaS apps and must be a workflow1–3 hours
  4. 4CrewAI / LangGraph / Agents SDKWhen you are shipping a product, not a personal cadenceHalf a day
  5. 5Meta-agents (agents that build agents)When the catalog is large enough to need a factoryDays

A (several named advisors, then a synthesizer) is the pattern behind the grant critic and the teaching program’s legal / values boards. Disagreement is a feature when the charter is shared.