# No-Code Agents in Claude Code

Based on insights by Tristen O'Brien (@Tristen_OBrien) in "Build Your FIRST AI Agent in Claude Code. (11 Minutes)." (published Nov 5, 2026, 10:54).

- Category: Coding
- Runs on: Claude
- Level: Intermediate
- Audience: both
- Evidence: Demonstrated (a vendor example shows the workflow; not run by us end to end)
Shared from [Tristen O'Brien \(@Tristen_OBrien\)](https://www.youtube.com/watch?v=ge2FmPeVsT8), who ran it on 2026-11-05. CrimsonBot hasn't tested it and can't promise it will work for you. Check each step against the source before you run it.
- Last updated: 2026-11-09
- Named sources: 2
- Canonical URL: https://crimsonbot.com/use-cases/no-code-agents-claude-code

> **Warning:** **Golden Rule of Implementation.** Never let the agent send email, post publicly, spend money, change logins, or delete files without an explicit human Approve. Configure connectors for drafts only until you trust the workflow. Stop the run if it asks for broader permissions than the job needs.

**Disclosure.** Personal agent-count and productivity framing in his chapters are the creator’s claims and were not independently measured here. Specific hour-saved or “100% delegation” style scores from secondary write-ups are omitted because they were not verified against the video narration in this build.

*This playbook is an educational synthesis intended to operationalize the insights presented in the source video. It is not affiliated with or endorsed by Tristen O’Brien or Anthropic. Watch the original for full context.*

## Risks

- Gmail/Workspace connector scopes may allow send or delete. Keep it drafts-only and revoke unused scopes.
- Prompt injection: task text, emails, or web pages the agent reads can carry instructions. Review before you approve actions.
- Sub-agents burn quota. Start small.
- Auto mode can edit or delete files. Approve the plan in Plan Mode first.
- The creator is paid by Anthropic (sponsored video).

## Steps

### 1. Executive summary and core framework

## 1. Executive Summary & Core Architectural Framework

Chat answers questions; an agent takes a goal, picks tools, and keeps going until the job is done—or until it hits a human gate. Tristen O’Brien’s walkthrough—regrouped here into four modules—shows non-engineers building that loop inside the Claude desktop app with Claude Code: a CLAUDE.md brain, Plan Mode with a brain-dump, sub-agents on a task list, the agent in your Claude account (usable from your phone, per creator), and drafts instead of live sends. The video is sponsored by Anthropic (per the description).

**The thesis.** The old way is you as the worker bee copying the model’s checklist into email and docs. O’Brien’s way hands the agent the list, runs helper agents in parallel, and keeps send/delete behind your click.

### 2. Core principles from the video

**Core principles from the video**

- **Goal in, steps out.** Agents plan and use tools; they are not only chat.
- **CLAUDE.md is durable memory.** Preferences and safety rules live in the workspace file.
- **Plan Mode before Auto.** Brain-dumps and clarifying questions beat one-line prompts.
- **Sub-agents parallelize work.** A coordinator splits the list so helpers run together.
- **Drafts, not silent sends.** External mail stages for your approval (he stresses it never sends without asking).

### 3. Module 1 — CLAUDE.md Brain & Safety Defaults

## 2. Playbook Modules

Covers workspace setup from Tristen O’Brien’s video.

**Problem statement & benchmark data.** Fresh chats forget your tone and rules. O’Brien starts from a workspace brain file so the agent reloads preferences every run. His channel framing includes a large personal agent count in the chapter titles (per creator); treat scale claims as his, not a benchmark you must hit on day one.

**Case study — from the video.** He sets up Claude Code in the Claude app (no terminal required for the beginner path he shows) and steers behavior with workspace instructions before building the to-do agent.

**Production prompt sequence**

**Governance control.** A bloated rules file dilutes itself. Prefer a few hard safety lines over hundreds of micro-rules.

```text
**Prompt 1 — Initialize CLAUDE.md**

Create a CLAUDE.md for my Claude Code workspace.
Domain: [PERSONAL OPS / WORK / OTHER]

Include:
- Tone: clear, short sentences, no corporate fluff
- Safety: never send email, publish, pay, or delete without my Approve
- Auto-OK: read files, draft locally, search
- Ask first: external sends, destructive file ops, new connectors

Do NOT connect Gmail yet. Do NOT send anything.

**Prompt 2 — Tighten permissions**

Revise CLAUDE.md so Plan Mode must finish and I must approve before Auto Mode edits files.
Add a short glossary: [TERMS].
Keep the file short enough to follow—no novel-length rules.
```

### 4. Module 2 — Plan Mode & Brain-Dump (human checkpoint)

Covers Plan Mode and the brain-dump (his "30-second rule" chapter) from Tristen O’Brien’s video.

**Problem statement & benchmark data.** Tiny typed prompts under-specify the job. He has you brain-dump the goal in Plan Mode, answer clarifying questions, then approve the plan before execution.

**Case study — from the video.** Building a to-do agent: brain-dump the outcome, let Claude question gaps, approve the plan, then allow implementation.

**Production prompt sequence**

**Governance control.** Skipping Plan Mode burns tokens and ships the wrong thing. Approve is a real gate, not theater.

```text
**Prompt 1 — Plan only**

Mode: Plan Mode only. Do NOT write code or create files yet.
Here is my brain-dump:
[PASTE]

1. Restate the goal in one sentence.
2. Ask up to 3 clarifying questions.
3. Outline files, UI, and tests you would build.
Wait for my Approve before Auto Mode.

**Prompt 2 — Acceptance checks**

Given plan [PASTE], list failure modes (rate limits, bad tool scopes, vague done-when).
Add measurable acceptance checks for each module.
Do NOT start build until I say Approve.
```

### 5. Module 3 — Sub-Agents on a Task List

Covers parallel helpers from Tristen O’Brien’s video.

**Problem statement & benchmark data.** One agent serializing every chore is slow and context-heavy. He demos giving a whole list so sub-agents work items together inside the Claude Code agent pattern.

**Case study — from the video.** A to-do list agent fans work across sub-agents, reports progress, and still blocks external send on confirmation.

**Production prompt sequence**

**Governance control.** Parallel runs use more rate limit/quota. Start small. Do not invent productivity scores.

```text
**Prompt 1 — Coordinator**

You are the coordinator agent.
Task list:
[PASTE ITEMS]

Spawn sub-agents only for items that need no external send.
For each item return: owner sub-agent, status, artifact path, needs-Approve (yes/no).
Do NOT email, tweet, purchase, or delete.
If an item needs outreach, mark Together/Needs human and draft only.

**Prompt 2 — Progress board**

Refresh the board: Done / In progress / Blocked / Needs Approve.
Estimate nothing as "hours saved" unless I supply timings.
Do NOT mark outreach Done until I confirm send.
```

### 6. Module 4 — Phone access and draft-only outreach (human checkpoint)

Module 4 — The agent in your Claude account (usable from your phone, per creator) + Draft-Only Outreach

Covers mobile/cloud use and Gmail-style drafting from Tristen O’Brien’s video.

**Problem statement & benchmark data.** He shows the agent living in your Claude account so you can drive it from your phone even when the desktop sleeps, and stresses that it should not send without asking—draft first.

**Case study — from the video.** The task agent used from the phone (per creator); communication paths stage for human send rather than silent dispatch.

**Production prompt sequence**

**Governance control.** Inspect the Drafts folder on first runs. Revoke send scopes you did not mean to grant. Anthropic sponsorship does not change your account security duties.

```text
**Prompt 1 — Mobile task-board brief**

Design a simple task-board view for the agent in your Claude account (usable from your phone):
columns Done / Doing / Together(Needs me).
Tapping a card shows status and draft links.
No send buttons that call a live mail API.
Do NOT enable messages.send.

**Prompt 2 — Draft-only mail**

For Together items: [LIST]
Create Gmail DRAFTS only (drafts.create). Never messages.send.
Subject + body per item. Return draft links.
Stop if the connector asks for send-wide scopes.
```

### 7. Master execution and deployment matrix

## 3. Master Execution & Deployment Matrix

**Build order:** CLAUDE.md safety → Plan Mode approve → sub-agent list → the agent in your Claude account (usable from your phone, per creator) with drafts-only mail.

```text
| Module / Focus | Operational Objective | Key Insight / Benchmark | Primary Prompt / Workflow Pattern | Governance & Safety Control |
|---|---|---|---|---|
| 1 CLAUDE.md brain | Durable prefs + safety | Workspace file steers every run | Init rules; ask-first list | Least privilege |
| 2 Plan Mode + brain-dump | Spec before build | Clarify then Approve | Brain-dump → questions → plan | No silent Auto |
| 3 Sub-agent list | Parallel chores | Coordinator + helpers | Fan-out board; no send | Quota awareness |
| 4 The agent in your Claude account (usable from your phone, per creator) + drafts | Phone access; safe outreach | Drafts before send (per creator) | Task board; drafts.create only | Human hits Send |
```

### 8. Credits and source references

## 4. Credits & Source References

**Source video:** Tristen O’Brien (@Tristen_OBrien), "Build Your FIRST AI Agent in Claude Code. (11 Minutes).," published Nov 5, 2026 (10:54). https://www.youtube.com/watch?v=ge2FmPeVsT8 — also https://youtu.be/ge2FmPeVsT8. Prompts are new writing based on his walkthrough. Description notes Anthropic sponsorship and a Claude Max referral (clau.de/TristenObrien)—creator/sponsor links, not CrimsonBot’s.

**Third parties named**

- Anthropic Claude desktop app and Claude Code
- The agent in your Claude account (usable from your phone, per creator)
- Google Workspace / Gmail-style connectors (as configured in product)

## Tools used

- [Claude Code](https://claude.com/product/claude-code): Coding agent for the terminal, IDE and desktop.

## Next

- [Our tool: Scrub a PDF locally before you paste it into AI →](https://crimsonbot.com/go/p/pdf-scrubber)
- [Get an email when the newsletter launches →](https://crimsonbot.com/#newsletter)